Jaipur, Rajasthan, India
2K followers
500+ connections
About
Articles by Harpreet Singh
Activity
2K followers
-
Harpreet Singh Sachdev shared this🚀 𝐈 𝐁𝐮𝐢𝐥𝐭 𝐌𝐲 𝐎𝐰𝐧 𝐋𝐨𝐜𝐚𝐥 𝐀𝐈 𝐒𝐭𝐚𝐜𝐤 𝐨𝐧 𝐚 𝐑𝐚𝐬𝐩𝐛𝐞𝐫𝐫𝐲 𝐏𝐢 (and Stopped Burning $$$ on Tokens) A few days ago, I hit a frustrating limit - 𝐦𝐲 𝐞𝐧𝐭𝐢𝐫𝐞 𝐦𝐨𝐧𝐭𝐡𝐥𝐲 𝐀𝐏𝐈 𝐛𝐮𝐝𝐠𝐞𝐭 𝐟𝐨𝐫 𝐩𝐞𝐫𝐬𝐨𝐧𝐚𝐥 𝐑&𝐃 𝐰𝐚𝐬 𝐛𝐮𝐫𝐧𝐞𝐝 𝐭𝐡𝐫𝐨𝐮𝐠𝐡 𝐢𝐧 𝐣𝐮𝐬𝐭 𝐭𝐰𝐨 𝐝𝐚𝐲𝐬. That’s when I decided : 𝑊ℎ𝑎𝑡 𝑖𝑓 𝐼 𝑐𝑜𝑢𝑙𝑑 𝑟𝑢𝑛 𝑒𝑣𝑒𝑟𝑦𝑡ℎ𝑖𝑛𝑔 𝑙𝑜𝑐𝑎𝑙𝑙𝑦? So I built a 𝐟𝐮𝐥𝐥𝐲 𝐬𝐞𝐥𝐟-𝐡𝐨𝐬𝐭𝐞𝐝 𝐀𝐈 𝐬𝐞𝐭𝐮𝐩 using a 𝐑𝐚𝐬𝐩𝐛𝐞𝐫𝐫𝐲 𝐏𝐢 - and it actually works. 💡𝐖𝐡𝐚𝐭 𝐈 𝐁𝐮𝐢𝐥𝐭 Instead of relying on expensive cloud APIs, I created a 𝐥𝐨𝐜𝐚𝐥 𝐋𝐋𝐌 𝐬𝐭𝐚𝐜𝐤 that runs entirely on my own hardware : • Hosted models using Ollama • Optimised inference with 𝐥𝐥𝐚𝐦𝐚.𝐜𝐩𝐩. • Applied quantization to run small-to-medium models efficiently on limited hardware. • Exposed everything over my local network via NGINX. • Built a custom Visual Studio Code extension (TypeScript-based, courtesy of Claude Code) to interact with my local models. • Designed a backend dashboard using Django with API key management • Containerized the entire stack using Docker with PostgreSQL 🧠 𝐖𝐡𝐲 𝐐𝐮𝐚𝐧𝐭𝐢𝐳𝐚𝐭𝐢𝐨𝐧 𝐖𝐚𝐬 𝐭𝐡𝐞 𝐆𝐚𝐦𝐞 𝐂𝐡𝐚𝐧𝐠𝐞𝐫 Running LLMs on a Raspberry Pi isn’t trivial. The breakthrough came from quantization - reducing model precision (e.g., from FP16 to 4-bit/8-bit), which: • Shrinks model size drastically • Reduces RAM usage • Makes CPU-only inference viable Without this, running models locally on such constrained hardware would be nearly impossible. 🏗️ 𝐇𝐢𝐠𝐡-𝐋𝐞𝐯𝐞𝐥 𝐀𝐫𝐜𝐡𝐢𝐭𝐞𝐜𝐭𝐮𝐫𝐞 VS Code Extension (TypeScript) ↓ NGINX (Auth + Reverse Proxy on Raspberry Pi) ↓ Ollama Runtime ↓ Quantized Models via llama.cpp ↓ Dockerized on Raspberry Pi ↓ PostgreSQL (Ollama Logs, API keys + RAG documents - managed by Django) 🔐 𝐖𝐡𝐚𝐭 𝐌𝐚𝐤𝐞𝐬 𝐓𝐡𝐢𝐬 𝐒𝐞𝐭𝐮𝐩 𝐏𝐨𝐰𝐞𝐫𝐟𝐮𝐥 • Zero token cost → No more surprise bills • Full control → My data never leaves my network • Custom workflows → Tailored exactly to my needs • Offline capability → Works even without internet ⚡ 𝐑𝐞𝐚𝐥𝐢𝐳𝐚𝐭𝐢𝐨𝐧 We often default to cloud APIs because they’re convenient. But convenience comes at a 𝐜𝐨𝐬𝐭 — 𝐥𝐢𝐭𝐞𝐫𝐚𝐥𝐥𝐲. In my case, I was burning through tokens faster than expected. Building this local system made me realize : 👉 Sometimes, 𝐞𝐧𝐠𝐢𝐧𝐞𝐞𝐫𝐢𝐧𝐠 𝐲𝐨𝐮𝐫 𝐨𝐰𝐧 𝐬𝐨𝐥𝐮𝐭𝐢𝐨𝐧 𝐢𝐬 𝐜𝐡𝐞𝐚𝐩𝐞𝐫, 𝐦𝐨𝐫𝐞 𝐬𝐜𝐚𝐥𝐚𝐛𝐥𝐞, 𝐚𝐧𝐝 𝐟𝐚𝐫 𝐦𝐨𝐫𝐞 𝐞𝐦𝐩𝐨𝐰𝐞𝐫𝐢𝐧𝐠. 🤝 𝐅𝐢𝐧𝐚𝐥 𝐓𝐡𝐨𝐮𝐠𝐡𝐭 This isn’t about replacing cloud AI - it’s about having an alternative. If you're experimenting, building tools, or learning - try going local at least once. 𝐘𝐨𝐮’𝐥𝐥 𝐥𝐞𝐚𝐫𝐧 𝐦𝐨𝐫𝐞 𝐭𝐡𝐚𝐧 𝐚𝐧𝐲 𝐀𝐏𝐈 𝐜𝐨𝐮𝐥𝐝 𝐞𝐯𝐞𝐫 𝐭𝐞𝐚𝐜𝐡 𝐲𝐨𝐮. #AI #LLM #LocalAI #RaspberryPi #Ollama #LlamaCpp #Quantization #RAG
-
Harpreet Singh Sachdev shared this𝐈𝐧𝐝𝐢𝐚’𝐬 𝐓𝐞𝐜𝐡 𝐇𝐢𝐫𝐢𝐧𝐠 𝐈𝐬 𝐐𝐮𝐢𝐞𝐭𝐥𝐲 𝐄𝐯𝐨𝐥𝐯𝐢𝐧𝐠 A recent Bank of America analysis highlights a shift in India’s tech hiring - from volume based hiring to high value, skill driven roles. AI is improving productivity, meaning companies may hire fewer people, but for more specialized work. The focus is clearly moving toward AI, data, and advanced engineering capabilities, while repetitive roles continue to decline. At the same time, India remains a key global tech hub due to its strong talent base and cost advantage. Over the next 5 years, hiring is likely to become more selective, skill-intensive, and outcome-driven. We may see slower headcount growth, but higher demand for niche expertise and problem-solving roles. The gap between high-skill and low-skill roles could widen significantly. 𝐑𝐞𝐚𝐝 𝐌𝐨𝐫𝐞 : https://lnkd.in/g6tMXQvU #IndiaTech #AI #FutureOfWork #UpskillingIndia’s tech hiring is changing: Bank of America says pivot toward high-value rolesIndia’s tech hiring is changing: Bank of America says pivot toward high-value roles
-
Harpreet Singh Sachdev shared this𝐂𝐨𝐦𝐦𝐨𝐧 𝐌𝐢𝐬𝐭𝐚𝐤𝐞𝐬 𝐖𝐡𝐢𝐥𝐞 𝐁𝐮𝐢𝐥𝐝𝐢𝐧𝐠 𝐚 𝐌𝐨𝐝𝐞𝐥 𝐂𝐨𝐧𝐭𝐞𝐱𝐭 𝐏𝐫𝐨𝐭𝐨𝐜𝐨𝐥 (𝐌𝐂𝐏) 𝐒𝐲𝐬𝐭𝐞𝐦 - 𝐚𝐧𝐝 𝐇𝐨𝐰 𝐭𝐨 𝐅𝐢𝐱 𝐓𝐡𝐞𝐦 I’ve noticed a pattern when teams start building systems using 𝐌𝐨𝐝𝐞𝐥 𝐂𝐨𝐧𝐭𝐞𝐱𝐭 𝐏𝐫𝐨𝐭𝐨𝐜𝐨𝐥 (𝐌𝐂𝐏). The idea is powerful - connect models with tools, context, and data. But the first implementation often goes wrong because of a few avoidable mistakes. 𝐌𝐢𝐬𝐭𝐚𝐤𝐞 𝟏: 𝐓𝐫𝐞𝐚𝐭𝐢𝐧𝐠 𝐌𝐂𝐏 𝐥𝐢𝐤𝐞 𝐚 𝐬𝐢𝐦𝐩𝐥𝐞 𝐀𝐏𝐈 𝐰𝐫𝐚𝐩𝐩𝐞𝐫 Many teams just expose APIs and call it MCP. But MCP is about structured context exchange, not just endpoints. For example, instead of simply giving an LLM access to a database API, provide structured context like schema, query intent, and constraints. This helps the model reason better. 𝐌𝐢𝐬𝐭𝐚𝐤𝐞 𝟐: 𝐏𝐨𝐨𝐫𝐥𝐲 𝐝𝐞𝐬𝐢𝐠𝐧𝐞𝐝 𝐜𝐨𝐧𝐭𝐞𝐱𝐭 𝐥𝐚𝐲𝐞𝐫𝐬 If the context sent to the model is messy or too large, performance drops quickly. Example : sending an entire CRM dataset when the task only needs customer purchase history and last interaction. 𝑇ℎ𝑒 𝑓𝑖𝑥 𝑖𝑠 𝑐𝑜𝑛𝑡𝑒𝑥𝑡 𝑓𝑖𝑙𝑡𝑒𝑟𝑖𝑛𝑔 𝑎𝑛𝑑 𝑟𝑒𝑡𝑟𝑖𝑒𝑣𝑎𝑙 𝑙𝑜𝑔𝑖𝑐. 𝐌𝐢𝐬𝐭𝐚𝐤𝐞 𝟑: 𝐈𝐠𝐧𝐨𝐫𝐢𝐧𝐠 𝐭𝐨𝐨𝐥 𝐛𝐨𝐮𝐧𝐝𝐚𝐫𝐢𝐞𝐬 Some implementations give the model too many tools with unclear roles. Imagine a system where the model can access SQL, email APIs, analytics tools, and file systems without defined rules - the model becomes inconsistent in choosing tools. 𝐼𝑛𝑠𝑡𝑒𝑎𝑑, 𝑑𝑒𝑠𝑖𝑔𝑛 𝑐𝑙𝑒𝑎𝑟 𝑡𝑜𝑜𝑙 𝑠𝑐ℎ𝑒𝑚𝑎𝑠 𝑎𝑛𝑑 𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑖𝑏𝑖𝑙𝑖𝑡𝑖𝑒𝑠. 𝐌𝐢𝐬𝐭𝐚𝐤𝐞 𝟒: 𝐍𝐨 𝐨𝐛𝐬𝐞𝐫𝐯𝐚𝐛𝐢𝐥𝐢𝐭𝐲 𝐨𝐟 𝐦𝐨𝐝𝐞𝐥-𝐭𝐨𝐨𝐥 𝐢𝐧𝐭𝐞𝐫𝐚𝐜𝐭𝐢𝐨𝐧𝐬 Teams often deploy MCP systems but don’t track how the model actually uses tools. Without logs, you can’t answer questions like : “Why did the model query the database five times for the same request?” 𝐌𝐢𝐬𝐭𝐚𝐤𝐞 𝟓: 𝐍𝐨𝐭 𝐝𝐞𝐬𝐢𝐠𝐧𝐢𝐧𝐠 𝐟𝐨𝐫 𝐢𝐭𝐞𝐫𝐚𝐭𝐢𝐯𝐞 𝐜𝐨𝐧𝐭𝐞𝐱𝐭 𝐮𝐩𝐝𝐚𝐭𝐞𝐬 MCP systems work best when context evolves during the workflow. For instance, a customer support agent might start with ticket data, then fetch billing history, and later access product documentation - context should grow logically as the task progresses. The biggest lesson : 𝐌𝐂𝐏 𝐢𝐬𝐧’𝐭 𝐣𝐮𝐬𝐭 𝐚𝐛𝐨𝐮𝐭 𝐜𝐨𝐧𝐧𝐞𝐜𝐭𝐢𝐧𝐠 𝐦𝐨𝐝𝐞𝐥𝐬 𝐭𝐨 𝐭𝐨𝐨𝐥𝐬 — 𝐢𝐭’𝐬 𝐚𝐛𝐨𝐮𝐭 𝐝𝐞𝐬𝐢𝐠𝐧𝐢𝐧𝐠 𝐭𝐡𝐞 𝐫𝐢𝐠𝐡𝐭 𝐟𝐥𝐨𝐰 𝐨𝐟 𝐜𝐨𝐧𝐭𝐞𝐱𝐭. When done well, MCP turns AI systems from simple chatbots into reliable AI agents. 🚀 #AIEngineering #ModelContextProtocol #AIAgents #LLMEngineering #AIArchitecture #AIInfrastructure #GenerativeAI #MachineLearning #AIProductDevelopment #TechLeadership
-
Harpreet Singh Sachdev shared thisIn LLMs, when a model is optimised purely for benchmark scores, it often learns to “game” the test rather than truly understand language. That’s reward hacking in disguise. We’re not that different. We celebrate 𝐥𝐞𝐚𝐝𝐞𝐫𝐛𝐨𝐚𝐫𝐝 𝐫𝐚𝐧𝐤𝐢𝐧𝐠𝐬, 𝐩𝐚𝐫𝐚𝐦𝐞𝐭𝐞𝐫 𝐜𝐨𝐮𝐧𝐭𝐬, and 𝐝𝐞𝐦𝐨 𝐯𝐢𝐫𝐚𝐥𝐢𝐭𝐲 - sometimes more than real user value. 𝐆𝐨𝐨𝐝𝐡𝐚𝐫𝐭’𝐬 𝐋𝐚𝐰 says it clearly : 𝙒𝙝𝙚𝙣 𝙩𝙝𝙚 𝙢𝙚𝙖𝙨𝙪𝙧𝙚 𝙗𝙚𝙘𝙤𝙢𝙚𝙨 𝙩𝙝𝙚 𝙩𝙖𝙧𝙜𝙚𝙩, 𝙞𝙩 𝙨𝙩𝙤𝙥𝙨 𝙗𝙚𝙞𝙣𝙜 𝙖 𝙜𝙤𝙤𝙙 𝙢𝙚𝙖𝙨𝙪𝙧𝙚. 🔎 𝐀 𝐒𝐦𝐚𝐥𝐥 𝐄𝐱𝐚𝐦𝐩𝐥𝐞 An LLM is fine-tuned to maximize benchmark accuracy. It memorizes common test patterns and overfits to evaluation sets. 𝐎𝐧 𝐩𝐚𝐩𝐞𝐫? State-of-the-art. 𝐈𝐧 𝐩𝐫𝐨𝐝𝐮𝐜𝐭𝐢𝐨𝐧? It hallucinates confidently and fails on messy, real-world queries. 𝐓𝐡𝐞 𝐬𝐜𝐨𝐫𝐞 𝐢𝐦𝐩𝐫𝐨𝐯𝐞𝐝. 𝐓𝐫𝐮𝐬𝐭 𝐝𝐢𝐝𝐧’𝐭. Benchmarks matter. Metrics matter. But usefulness, reliability, and real-world impact matter more. The real question is : Are we training models 𝐭𝐨 𝐰𝐢𝐧 𝐭𝐞𝐬𝐭𝐬 - 𝐨𝐫 𝐭𝐨 𝐬𝐨𝐥𝐯𝐞 𝐩𝐫𝐨𝐛𝐥𝐞𝐦𝐬? #LLM #ArtificialIntelligence #MachineLearning #GenerativeAI #AIResearch #ResponsibleAI #GoodhartsLaw #AIEvaluation #AILeadership #TechStrategy
-
Harpreet Singh Sachdev shared this𝐀𝐈 𝐢𝐧 𝐏𝐫𝐨𝐝𝐮𝐜𝐭𝐢𝐨𝐧: 𝐀 𝐋𝐞𝐚𝐝𝐞𝐫𝐬𝐡𝐢𝐩 𝐖𝐚𝐤𝐞-𝐔𝐩 𝐂𝐚𝐥𝐥 𝐨𝐧 𝐀𝐜𝐜𝐞𝐬𝐬 𝐂𝐨𝐧𝐭𝐫𝐨𝐥 Recent reporting by India Today and coverage from Financial Times spotlighted a disruption involving Amazon Web Services and an internal AI coding agent known as “Kiro.” The reported incident referenced a 𝟏𝟑-𝐡𝐨𝐮𝐫 𝐝𝐢𝐬𝐫𝐮𝐩𝐭𝐢𝐨𝐧, where the agent allegedly chose to “𝐝𝐞𝐥𝐞𝐭𝐞 𝐚𝐧𝐝 𝐫𝐞𝐜𝐫𝐞𝐚𝐭𝐞 𝐭𝐡𝐞 𝐞𝐧𝐯𝐢𝐫𝐨𝐧𝐦𝐞𝐧𝐭.” AWS stated the event was limited in scope, affecting AWS Cost Explorer in a single Mainland China region, and attributed the root cause to misconfigured access controls, not runaway AI autonomy. Here’s the takeaway: This is not a debate about “AI error” vs “user error.” It’s about 𝐠𝐨𝐯𝐞𝐫𝐧𝐚𝐧𝐜𝐞, 𝐚𝐜𝐜𝐞𝐬𝐬 𝐝𝐞𝐬𝐢𝐠𝐧, and 𝐨𝐩𝐞𝐫𝐚𝐭𝐢𝐨𝐧𝐚𝐥 𝐠𝐮𝐚𝐫𝐝𝐫𝐚𝐢𝐥𝐬. When AI agents are granted production-level autonomy, a permissions oversight is no longer minor - it becomes a machine-speed failure. AI operates instantly. Your controls must be stronger than ever. For leadership teams deploying AI into engineering workflows, this is the new non-negotiable checklist: • Least-privilege access by default • Clear separation of duties • Explicit approval for destructive actions • Strict blast-radius containment • Mandatory peer review for infra changes • Immutable logs and audit trails • Tested rollback and recovery paths AI doesn’t eliminate responsibility. It amplifies consequences. In the agent era, access control is no longer just a security setting. 𝐈𝐭’𝐬 𝐞𝐱𝐞𝐜𝐮𝐭𝐢𝐯𝐞 𝐫𝐢𝐬𝐤 𝐦𝐚𝐧𝐚𝐠𝐞𝐦𝐞𝐧𝐭. #AI #CyberSecurity #Leadership #CloudComputing #AWS #RiskManagement #DevOps #Governance
-
Harpreet Singh Sachdev shared this𝐖𝐡𝐞𝐧 𝐀𝐈 𝐈𝐬 𝐂𝐨𝐧𝐟𝐢𝐝𝐞𝐧𝐭… 𝐛𝐮𝐭 𝐖𝐫𝐨𝐧𝐠 I came across a 𝐑𝐞𝐝𝐝𝐢𝐭 story this week that genuinely made me pause. A company had been using an AI agent to answer leadership questions about their metrics. It was 𝐟𝐚𝐬𝐭, 𝐚𝐫𝐭𝐢𝐜𝐮𝐥𝐚𝐭𝐞, 𝐚𝐧𝐝 𝐢𝐧𝐜𝐫𝐞𝐝𝐢𝐛𝐥𝐲 𝐜𝐨𝐧𝐟𝐢𝐝𝐞𝐧𝐭. 𝐄𝐯𝐞𝐫𝐲𝐨𝐧𝐞 𝐭𝐫𝐮𝐬𝐭𝐞𝐝 𝐢𝐭. Three months later, they discovered it had been 𝐟𝐚𝐛𝐫𝐢𝐜𝐚𝐭𝐢𝐧𝐠 𝐚𝐧𝐚𝐥𝐲𝐭𝐢𝐜𝐬 𝐝𝐚𝐭𝐚. ✦ Some numbers came from the wrong time periods. ✦ Some mixed up products. ✦ Some were simply made up — but explained convincingly enough that no one questioned them. ➢ Decisions were made. ➢ Board decks were presented. ➢ Territory strategies were adjusted. The issue wasn’t just that the AI made mistakes. It’s that it made them confidently. Here’s the real takeaway: ➳ 𝐆𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐯𝐞 𝐀𝐈 𝐢𝐬 𝐧𝐨𝐭 𝐚 𝐬𝐲𝐬𝐭𝐞𝐦 𝐨𝐟 𝐫𝐞𝐜𝐨𝐫𝐝. ➳ 𝐂𝐨𝐧𝐟𝐢𝐝𝐞𝐧𝐜𝐞 𝐝𝐨𝐞𝐬 𝐧𝐨𝐭 𝐞𝐪𝐮𝐚𝐥 𝐜𝐨𝐫𝐫𝐞𝐜𝐭𝐧𝐞𝐬𝐬. ➳ 𝐒𝐩𝐞𝐞𝐝 𝐰𝐢𝐭𝐡𝐨𝐮𝐭 𝐠𝐨𝐯𝐞𝐫𝐧𝐚𝐧𝐜𝐞 𝐜𝐫𝐞𝐚𝐭𝐞𝐬 𝐫𝐢𝐬𝐤. As we embed AI deeper into analytics and decision-making, validation can’t be optional — it has to be designed in. #AI #Analytics #DataGovernance #ResponsibleAI #Leadership
-
Harpreet Singh Sachdev shared this𝐖𝐡𝐚𝐭 𝐈𝐟 𝐘𝐨𝐮𝐫 𝐀𝐜𝐭𝐢𝐯𝐚𝐭𝐢𝐨𝐧 𝐅𝐮𝐧𝐜𝐭𝐢𝐨𝐧 𝐈𝐬 𝐐𝐮𝐢𝐞𝐭𝐥𝐲 𝐒𝐭𝐞𝐞𝐫𝐢𝐧𝐠 𝐎𝐩𝐭𝐢𝐦𝐢𝐳𝐚𝐭𝐢𝐨𝐧? We spend a lot of time debating architectures - Transformers vs CNNs, depth vs width, scaling laws, attention variants. But one of the most overlooked design choices is the 𝐚𝐜𝐭𝐢𝐯𝐚𝐭𝐢𝐨𝐧 𝐟𝐮𝐧𝐜𝐭𝐢𝐨𝐧. 𝐑𝐞𝐋𝐔 became the default because it solved 𝑣𝑎𝑛𝑖𝑠ℎ𝑖𝑛𝑔 𝑔𝑟𝑎𝑑𝑖𝑒𝑛𝑡𝑠 and made deep networks trainable at scale. It’s simple, efficient, and works well in most cases. But it also zeroes out all negative inputs. In domains like audio modelling, where signals naturally oscillate around zero, this can discard meaningful information. Half the waveform can effectively be muted. In several deep speech and audio enhancement models, switching from 𝐑𝐞𝐋𝐔 𝐭𝐨 𝐄𝐋𝐔 has led to more stable training at higher learning rates. Why? Because ELU allows controlled negative outputs, helping keep activations centered around zero and improving gradient flow across many layers. The improvement per layer is small. But across 40–60 layers, those small differences compound and influence convergence behavior. Activation functions don’t just add non-linearity. They shape 𝐬𝐢𝐠𝐧𝐚𝐥 𝐝𝐢𝐬𝐭𝐫𝐢𝐛𝐮𝐭𝐢𝐨𝐧, 𝐠𝐫𝐚𝐝𝐢𝐞𝐧𝐭 𝐝𝐲𝐧𝐚𝐦𝐢𝐜𝐬, and 𝐮𝐥𝐭𝐢𝐦𝐚𝐭𝐞𝐥𝐲 𝐭𝐡𝐞 𝐨𝐩𝐭𝐢𝐦𝐢𝐳𝐚𝐭𝐢𝐨𝐧 𝐩𝐚𝐭𝐡 𝐲𝐨𝐮𝐫 𝐦𝐨𝐝𝐞𝐥 𝐟𝐨𝐥𝐥𝐨𝐰𝐬. Sometimes, performance gains aren’t about bigger models - they’re about better geometry. #DeepLearning #MachineLearning #NeuralNetworks #AI #ModelOptimization #ArtificialIntelligence
-
Harpreet Singh Sachdev reacted on this𝐖𝐡𝐞𝐧 𝐃𝐢𝐝 𝐖𝐞 𝐅𝐨𝐫𝐠𝐞𝐭 𝐓𝐡𝐚𝐭 𝐒𝐢𝐦𝐩𝐥𝐞 𝐒𝐨𝐥𝐮𝐭𝐢𝐨𝐧𝐬 𝐄𝐱𝐢𝐬𝐭? A conversation I had yesterday : 𝐅𝐫𝐢𝐞𝐧𝐝 : We need AI for our project. 𝐌𝐞 : What's the problem? 𝐅𝐫𝐢𝐞𝐧𝐝 : Excel keeps crashing with our sales data. 𝐏𝐫𝐨𝐩𝐨𝐬𝐞𝐝 𝐬𝐨𝐥𝐮𝐭𝐢𝐨𝐧 : Deep learning model, cloud GPUs, real-time dashboard. 𝐀𝐜𝐭𝐮𝐚𝐥 𝐩𝐫𝐨𝐛𝐥𝐞𝐦 : 500,000 rows in one sheet with 47 pivot tables running simultaneously. 𝐑𝐞𝐚𝐥 𝐬𝐨𝐥𝐮𝐭𝐢𝐨𝐧 : Basic data aggregation + a simple linear regression model for forecasting. 𝐏𝐑𝐎𝐁𝐋𝐄𝐌 𝐒𝐎𝐋𝐕𝐄𝐃 𝐈𝐍 𝟐 𝐇𝐎𝐔𝐑𝐒. 𝐓𝐡𝐞 𝐜𝐮𝐫𝐫𝐞𝐧𝐭 𝐬𝐭𝐚𝐭𝐞 - 50 data points → "Let's use deep learning" - Basic if-else logic → "Can we fine-tune an LLM?" - Simple pattern recognition → "Need a neural network?" 𝐑𝐞𝐚𝐥𝐢𝐭𝐲 𝐜𝐡𝐞𝐜𝐤 : Not everything needs transformers, GPU clusters, or 20-layer architectures. Sometimes a mean, median, or basic regression tells you exactly what you need to know. 𝐂𝐥𝐞𝐚𝐧 𝐝𝐚𝐭𝐚 + 𝐬𝐢𝐦𝐩𝐥𝐞 𝐦𝐨𝐝𝐞𝐥 + 𝐜𝐨𝐦𝐦𝐨𝐧 𝐬𝐞𝐧𝐬𝐞 often beats complexity. We've created an environment where solutions aren't "impressive" unless they mention AI. 𝐓𝐡𝐞 𝐛𝐞𝐬𝐭 𝐝𝐚𝐭𝐚 𝐬𝐜𝐢𝐞𝐧𝐜𝐞? 𝐊𝐧𝐨𝐰𝐢𝐧𝐠 𝐰𝐡𝐞𝐧 𝐍𝐎𝐓 𝐭𝐨 𝐮𝐬𝐞 𝐭𝐡𝐞 𝐟𝐚𝐧𝐜𝐲 𝐬𝐭𝐮𝐟𝐟. #DataScience #MachineLearning #AI #RealityCheck
-
Harpreet Singh Sachdev reacted on this𝐖𝐡𝐞𝐧 𝐃𝐢𝐝 𝐖𝐞 𝐅𝐨𝐫𝐠𝐞𝐭 𝐓𝐡𝐚𝐭 𝐒𝐢𝐦𝐩𝐥𝐞 𝐒𝐨𝐥𝐮𝐭𝐢𝐨𝐧𝐬 𝐄𝐱𝐢𝐬𝐭? A conversation I had yesterday : 𝐅𝐫𝐢𝐞𝐧𝐝 : We need AI for our project. 𝐌𝐞 : What's the problem? 𝐅𝐫𝐢𝐞𝐧𝐝 : Excel keeps crashing with our sales data. 𝐏𝐫𝐨𝐩𝐨𝐬𝐞𝐝 𝐬𝐨𝐥𝐮𝐭𝐢𝐨𝐧 : Deep learning model, cloud GPUs, real-time dashboard. 𝐀𝐜𝐭𝐮𝐚𝐥 𝐩𝐫𝐨𝐛𝐥𝐞𝐦 : 500,000 rows in one sheet with 47 pivot tables running simultaneously. 𝐑𝐞𝐚𝐥 𝐬𝐨𝐥𝐮𝐭𝐢𝐨𝐧 : Basic data aggregation + a simple linear regression model for forecasting. 𝐏𝐑𝐎𝐁𝐋𝐄𝐌 𝐒𝐎𝐋𝐕𝐄𝐃 𝐈𝐍 𝟐 𝐇𝐎𝐔𝐑𝐒. 𝐓𝐡𝐞 𝐜𝐮𝐫𝐫𝐞𝐧𝐭 𝐬𝐭𝐚𝐭𝐞 - 50 data points → "Let's use deep learning" - Basic if-else logic → "Can we fine-tune an LLM?" - Simple pattern recognition → "Need a neural network?" 𝐑𝐞𝐚𝐥𝐢𝐭𝐲 𝐜𝐡𝐞𝐜𝐤 : Not everything needs transformers, GPU clusters, or 20-layer architectures. Sometimes a mean, median, or basic regression tells you exactly what you need to know. 𝐂𝐥𝐞𝐚𝐧 𝐝𝐚𝐭𝐚 + 𝐬𝐢𝐦𝐩𝐥𝐞 𝐦𝐨𝐝𝐞𝐥 + 𝐜𝐨𝐦𝐦𝐨𝐧 𝐬𝐞𝐧𝐬𝐞 often beats complexity. We've created an environment where solutions aren't "impressive" unless they mention AI. 𝐓𝐡𝐞 𝐛𝐞𝐬𝐭 𝐝𝐚𝐭𝐚 𝐬𝐜𝐢𝐞𝐧𝐜𝐞? 𝐊𝐧𝐨𝐰𝐢𝐧𝐠 𝐰𝐡𝐞𝐧 𝐍𝐎𝐓 𝐭𝐨 𝐮𝐬𝐞 𝐭𝐡𝐞 𝐟𝐚𝐧𝐜𝐲 𝐬𝐭𝐮𝐟𝐟. #DataScience #MachineLearning #AI #RealityCheck
-
Harpreet Singh Sachdev liked thisHarpreet Singh Sachdev liked thisByteDance's new open-source AI controls your entire computer. 100% local. Most desktop agents today rely on cloud APIs. Your screen, clicks, and files all leave the machine. ByteDance just open-sourced TARS, a multimodal agent stack that runs fully on your computer. It hit 29k stars on GitHub. The system uses a vision-language model to see pixels directly. It then drives your mouse and keyboard through plain English instructions. No screen reader. No accessibility hooks. Just raw visual grounding across any app on Windows, MacOS, or browser. Screenshots, recognition, and execution all stay local. Two pieces ship together: 1. CLI plus Web UI frontend 2. Native desktop GUI operator The kernel is built on MCP, plugging into real-world tools without custom glue. Hybrid browser control switches between visual targeting and DOM parsing mid-task. One prompt can book flights, edit spreadsheets, or scrape dashboards offline. Link in comments. ↓ Check out AlphaSignal.ai to get a daily summary of top models, repos, and papers in AI. Read by 280,000+ devs.
-
Harpreet Singh Sachdev liked thisHarpreet Singh Sachdev liked thisI stopped wasting compute on tokens like "the" and "a". (I saved 40% inference compute without losing a single accuracy point) For years, we've treated every token as equal. Quantization, pruning, sparse attention, all applied uniformly across the entire sequence. We assumed efficiency meant compressing everything. But here's the truth: In practice, token difficulty varies dramatically. A period is trivial. The final step of a complex math proof is not. Yet our models spend the same compute on both. That bothered me. A lot. Then I found *SOL* (Compute Where it Counts) this week. The Lesson: The path to efficiency isn't compressing everything equally. It's spending compute *where* it actually makes a difference. (some fun facts) Every token deserves the same computation budget. Easy token, hard token, same number of FLOPs. We've all accepted this as the cost of doing business. 1. The Problem: Under uniform compression, the 7.3% of accuracy we lose on MMLU isn't random, it's concentrated on the tokens that needed that compute most. We're starving the hard steps to feed the easy ones. 2. The SOL Solution: Pair a *frozen base LLM* with a *lightweight policy network* that reads the LLM's hidden state and selects the most efficient compute action for each token, on the fly, without ever touching the original weights. 3. The Mathematical Insight: The policy jointly optimizes three control knobs—(i) token-level attention sparsity, (ii) structured activation pruning in the MLP, and (iii) activation quantization bit-width. Because the base model remains unchanged, the policy learns to route efficiently without degrading the core capability. The Implementation: I integrated the `akhauriyash/SOL` repo (official code) into an existing inference pipeline. The policy training uses group-relative optimization on teacher-forced episodes, sampling compute schedules and optimizing a quality-efficiency reward. The result of solving compute allocation as a learning problem: - Dynamic budget allocation: Each token gets a unique compute schedule. - Quality boost: SOL improves MMLU accuracy by *up to 7.3% over static uniform strategies*. - Better Pareto frontier: Across all model variants, SOL consistently dominates fixed allocation, achieving higher accuracy at the same compute budget or lower compute at the same accuracy. (I think of it as) Stop assuming all tokens are born equal. Some need more compute than others. Let the model tell you which ones! #MachineLearning #LLM #Optimization #ICML2026
-
Harpreet Singh Sachdev liked thisHarpreet Singh Sachdev liked thisHugging Face open-sourced a CLI agent that runs full research loop of an ML engineer. You give it a prompt. It reads papers and walks citations. Then it pulls datasets and writes training scripts. It launches GPU jobs and iterates until something works. On scientific reasoning, it trained Qwen3-1.7B and pushed GPQA from 10% to 32%. Claude Code only reached 22.99%. For healthcare, it judged public datasets too weak. So it generated 1,100 synthetic samples and upsampled 50x. It beat Codex on HealthBench by 60%. On competitive math, rewards collapsed mid-run. The agent ran ablations and recovered the training. Three things make it work: 1. Reads papers and citation graphs 2. Reformats messy data before training 3. Diagnoses failed evals autonomously Ships as a CLI and a web app you can run from your phone. Link in comments. ↓ Check out AlphaSignal.ai to get a daily summary of top models, repos, and papers in AI. Read by 280,000+ devs.
-
Harpreet Singh Sachdev liked thisHarpreet Singh Sachdev liked thisQuantum Computing Reaches Major Milestone in Biological Simulation Interesting Engineering reports that researchers from IBM, the Cleveland Clinic, and Japan’s RIKEN institute achieved a major breakthrough by simulating a protein complex containing 12,635 atoms using quantum computing hardware. The accomplishment represents the largest biologically meaningful molecular simulation ever performed with quantum systems and marks an important advance toward practical biomedical and pharmaceutical applications. The research combined quantum processors with some of the world’s most powerful classical supercomputers in a hybrid computational architecture. Scientists modeled trypsin, a digestive enzyme commonly used in pharmaceutical research, along with another protein-ligand system. Classical systems divided the molecules into computational fragments while IBM’s 156-qubit Heron quantum processors calculated the quantum-mechanical interactions inside each section. The quantum calculations used as many as 94 qubits and nearly 6,000 operations per simulation segment. The resulting data was reconstructed using advanced classical systems including Japan’s Fugaku supercomputer and the Miyabi-G platform. Researchers stated the simulation achieved roughly 210 times greater accuracy than certain conventional approximation methods while scaling approximately 40 times larger than their previous milestone achieved only six months earlier. The broader significance lies in drug discovery and molecular science. Biological interactions inside proteins and pharmaceutical compounds are governed by quantum mechanics, making them extremely difficult for traditional computers to simulate accurately at large scales. Hybrid quantum-classical systems could eventually accelerate pharmaceutical development, molecular analysis, and precision medicine research. Key takeaways from the article suggest quantum computing is steadily moving from theoretical promise toward practical scientific utility. While fully fault-tolerant quantum systems are still developing, this breakthrough demonstrates that hybrid quantum architectures are already beginning to solve meaningful real-world research problems in chemistry, medicine, and advanced materials science. I share daily insights with tens of thousands followers across defense, tech, and policy. If this topic resonates, I invite you to connect and continue the conversation. Keith King https://lnkd.in/gHPvUttw
-
Harpreet Singh Sachdev liked thisHarpreet Singh Sachdev liked this#Hiring a Senior Product Designer (4–10 yrs) in Hyderabad at ASBL. We’re re-imagining how high-rises are built and redesigning construction site planning and on-ground ops. Looking for someone who has designed for field-heavy environments (like Zomato riders, Amazon warehouse ops, or Blue Dart logistics) You’ll work closely with Dilip Merugu & Abhishek Sanjapogu. Do reach out if this is for you.
-
Harpreet Singh Sachdev liked thisHarpreet Singh Sachdev liked thisYou can cut Claude Code's tool calls by 94% with one command. When Claude Code explores a repo, it spawns agents to scan files. Each scan burns tokens on every tool call. CodeGraph is an open-source MCP server that fixes this. It pre-indexes your codebase into a local knowledge graph upfront. Instead of scanning, the agent queries the graph directly. One call returns entry points, related symbols, and code snippets. Tested across six real-world projects, the results showed: > 92% fewer tool calls average > 71% faster exploration overall > 94% drop on a TypeScript repo > 96% drop on a Java codebase Tree-sitter parses source into syntax trees. SQLite stores symbols, callers, and edges with full-text search. A file watcher keeps everything fresh as you code. It supports 19+ languages and runs fully on your machine. Setup is one npx command. No API keys are required. Link in comments. ↓ Check out AlphaSignal.ai to get a daily summary of top models, repos, and papers in AI. Read by 280,000+ devs.
-
Harpreet Singh Sachdev liked thisHarpreet Singh Sachdev liked thisIf you're reading something today, let it be these 12 books on AI: 1. AI Engineering — How production systems actually get built. https://lnkd.in/g5Uin_PY 2. Designing ML Systems — Why real AI fails and how to prevent it. https://lnkd.in/gnJU_C9w 3. Prompt Engineering for LLMs — Structured techniques, not guesswork. https://lnkd.in/gTsxp_NH 4. Hands-On LLMs — Transformers and attention, finally explained clearly. https://lnkd.in/gKFGFjNe 5. LLM Engineer's Handbook — From fine-tuning to shipping. https://lnkd.in/gWkxZE77 6. Building LLMs From Scratch — Understand every layer from the ground up. https://lnkd.in/gHpNMh2y 7. Practical MLOps — Keeping models alive in the real world. https://lnkd.in/gS9NU2pn 8. Building LLMs for Production — The cost, speed, and reliability tradeoffs no one talks about. https://lnkd.in/gmiNQ4Jt 9. Foundations of Machine Learning — Solid theory before anything else. https://lnkd.in/d8QyxUiz 10. Deep Learning — The book that defined the field. https://lnkd.in/gi82Vn6y 11. Mathematics for ML — Just the math that matters. https://mml-book.github.io 12. Hands-On ML — Where concepts become working code. https://lnkd.in/g8kQsV8G
-
Harpreet Singh Sachdev liked thisHarpreet Singh Sachdev liked thisCPU vs GPU vs TPU vs NPU vs LPU 5 hardware architectures power AI today. Each one makes a fundamentally different tradeoff between flexibility, parallelism, and memory access. → CPU Built for general-purpose computing. A few powerful cores handle complex logic, branching, and system-level tasks. Deep cache hierarchies, off-chip DRAM. Great for operating systems, databases, and decision-heavy code. → GPU Instead of a few powerful cores, GPUs distribute work across thousands of smaller cores, each executing the same instruction on different data. This is why GPUs dominate AI training. The parallelism maps directly to the math neural networks need. → TPU Goes one step further with specialisation. The core compute unit is a grid of multiply-accumulate (MAC) units where data flows through in a wave pattern. Weights enter from one side, activations from the other, and partial results propagate without going back to memory each time. → NPU The edge-optimised variant. Built around a Neural Compute Engine packed with MAC arrays and on-chip SRAM. But instead of high-bandwidth memory (HBM), NPUs use low-power system memory. Smartphones. Wearables. IoT devices. Apple's Neural Engine and Intel's NPU follow this pattern. → LPU The newest entrant is by Groq. The architecture removes off-chip memory from the critical path entirely. All weight storage lives in on-chip SRAM. Execution is fully deterministic and compiler-scheduled. Zero cache misses. Zero runtime scheduling overhead. The visual below maps the internal architecture of all five, side by side 👇 Which one will define the next decade of AI? #AI #MachineLearning #DeepLearning
Experience & Education
View Harpreet Singh’s full experience
See their title, tenure and more.
or
Licenses & Certifications
Publications
-
The "Expert" Button: Can AI Really Magic Up Knowledge?
Harpreet Singh Sachdev
See publicationEver tried the "writing expert mode" with AI and got predictable results? Here's the scoop: AI, like ChatGPT, relies on prior training, so it won't surprise you with new info. But, here's the real magic: Treat AI as your creative buddy! Think of it as a friend who suggests and refines, not a mind-controller.
You're the artist with unique ideas; AI adds the knowledge blend.
1. Architect Your Masterpiece: Step by Step : Plan with AI. Align your vision step by step.
2. Clarify the…Ever tried the "writing expert mode" with AI and got predictable results? Here's the scoop: AI, like ChatGPT, relies on prior training, so it won't surprise you with new info. But, here's the real magic: Treat AI as your creative buddy! Think of it as a friend who suggests and refines, not a mind-controller.
You're the artist with unique ideas; AI adds the knowledge blend.
1. Architect Your Masterpiece: Step by Step : Plan with AI. Align your vision step by step.
2. Clarify the Fog: Seek Understanding Before Speed : Communicate clearly. Refine AI's grasp of your needs.
3. Fill the Knowledge Gaps: Build on Solid Ground : Source facts wisely. Ensure unbiased, well-rounded content.
4. Take A Breath, Refine the Gem : Pause for accuracy. Let AI align with your request.
Discover detailed insights in the article that discusses leveraging AI by refining instructions, transforming it from a mere content generator or task performer into an advisor. Now, it becomes someone who rectifies, suggests ideas, and fine-tunes tasks. Tips above can help you wield AI's knowledge base alongside your creativity to create pure magic! -
AI Drift In Retrieval Augmented Generation and ways to control it
Harpreet Singh Sachdev
See publicationThis articles talks about AI Drift, the gradual decline in AI performance, is examined within the context of Retrieval Augmented Generation (RAG). Analogous to a chameleon's color-changing abilities, RAG's AI responses may deviate over time due to changes in the retrieval source, biased data, and shifts in ranking and generative algorithms. Three types of drift are identified: Data Drift, Model Drift, and Interaction Drift. Data Drift results from outdated retrieval sources, while Model Drift…
This articles talks about AI Drift, the gradual decline in AI performance, is examined within the context of Retrieval Augmented Generation (RAG). Analogous to a chameleon's color-changing abilities, RAG's AI responses may deviate over time due to changes in the retrieval source, biased data, and shifts in ranking and generative algorithms. Three types of drift are identified: Data Drift, Model Drift, and Interaction Drift. Data Drift results from outdated retrieval sources, while Model Drift involves changes in ranking and generative models. Interaction Drift is influenced by user feedback and evolving expectations. Controlling AI drift in RAG demands a comprehensive strategy involving data hygiene, bias mitigation, model monitoring, and human and user feedback integration. This proactive, ongoing approach ensures RAG's reliability in storytelling and information retrieval.
-
Large Language Models (LLMs): Understanding and Optimizing for Programmatic Use
Harpreet Singh Sachdev
See publicationThe emergence of LLMs like Chatgpt , Google Gemini, Google Bard, and LlamaIndex has revolutionized the field of artificial intelligence. These powerful models, trained on vast datasets, excel at understanding, summarizing, generating, and predicting text content. For AI enthusiasts, the ability to interact with these models programmatically through Python opens exciting possibilities. However, navigating the intricacies of parameters can be challenging. This guide focuses on three key…
The emergence of LLMs like Chatgpt , Google Gemini, Google Bard, and LlamaIndex has revolutionized the field of artificial intelligence. These powerful models, trained on vast datasets, excel at understanding, summarizing, generating, and predicting text content. For AI enthusiasts, the ability to interact with these models programmatically through Python opens exciting possibilities. However, navigating the intricacies of parameters can be challenging. This guide focuses on three key parameters (temperature, top_k, top_p) that can significantly impact the quality and creativity of your LLM outputs
-
Sim Swap Fraud
Harpreet Singh Sachdev
See publicationIn this era of technological advancement, the double-edged nature of technology is evident through the emergence of cyber threats like SIM swap attacks. This sophisticated form of high-tech fraud involves manipulating individuals by momentarily disrupting their phone network, followed by a call from a purported representative of their cell phone company. The unsuspecting victim is then coerced into pressing a button, leading to the compromise of their SIM card and subsequent unauthorized access…
In this era of technological advancement, the double-edged nature of technology is evident through the emergence of cyber threats like SIM swap attacks. This sophisticated form of high-tech fraud involves manipulating individuals by momentarily disrupting their phone network, followed by a call from a purported representative of their cell phone company. The unsuspecting victim is then coerced into pressing a button, leading to the compromise of their SIM card and subsequent unauthorized access to sensitive information, enabling hackers to conduct fraudulent bank transactions. The hackers leverage extensive personal information obtained through online searches, making it crucial for individuals to minimize their digital footprint and exercise caution with unknown calls and messages. Safeguard measures include employing strong account security practices, utilizing authentication apps, and limiting the disclosure of personal details online. Awareness and vigilance serve as key components in mitigating the risks associated with this type of cyber attack.
-
Machine Learning vs Deep Learning
Harpreet Singh Sachdev
See publicationThis article discusses the dilemma faced in choosing between machine learning and deep learning algorithms when addressing data science problems. It emphasizes that machine learning is suitable for datasets with fewer features, where data parsing and pattern learning are relatively straightforward, using examples like recommendation algorithms in streaming services. On the other hand, deep learning, a subset of machine learning, is recommended for scenarios with numerous features or…
This article discusses the dilemma faced in choosing between machine learning and deep learning algorithms when addressing data science problems. It emphasizes that machine learning is suitable for datasets with fewer features, where data parsing and pattern learning are relatively straightforward, using examples like recommendation algorithms in streaming services. On the other hand, deep learning, a subset of machine learning, is recommended for scenarios with numerous features or multidimensional data, particularly when extracting features becomes tedious. The article illustrates this with an example of a cat vs dog image classifier, highlighting the challenge of dealing with a vast number of features and the potential use of convolutional neural networks (CNN) in deep learning to automatically extract relevant features. It also discusses the role of dimensionality reduction techniques like PCA and concludes that deep learning becomes beneficial as data volume increases and feature extraction becomes challenging even after applying dimensionality reduction. The key distinction lies in the autonomy of deep learning models in assessing prediction accuracy without manual intervention.
-
Choosing number of Hidden Layers and number of hidden neurons in Neural Networks
Harpreet Singh Sachdev
See publicationThis article provides a pragmatic approach to the often challenging task of determining the architecture of neural networks during deep learning model development. It emphasizes the importance of considering the complexity and dimensionality of the data when deciding on the number of hidden layers, suggesting that linearly separable data may not require any hidden layers, while less complex data could benefit from 1 to 2 hidden layers, and more complex data may necessitate 3 to 5 hidden layers…
This article provides a pragmatic approach to the often challenging task of determining the architecture of neural networks during deep learning model development. It emphasizes the importance of considering the complexity and dimensionality of the data when deciding on the number of hidden layers, suggesting that linearly separable data may not require any hidden layers, while less complex data could benefit from 1 to 2 hidden layers, and more complex data may necessitate 3 to 5 hidden layers for optimal solutions. The article also offers guidelines for selecting the number of nodes in each hidden layer, advocating for a decreasing pattern to facilitate pattern and feature extraction. It acknowledges the flexibility of these rules, allowing for adjustments based on specific use cases and problem statements, recognizing that factors such as overfitting may be influenced by the chosen architecture. Overall, the article provides a concise set of rules to streamline the neural network design process and mitigate the challenges of tuning hidden layers and nodes.
-
Fake News Detection using Deep Learning
Harpreet Singh Sachdev
See publicationFake news is a major concern in our society right now. Fake News is any news that is either factually wrong, misrepresents the facts, and that spreads virally (or maybe to a targeted audience). We believe what we saw and sometimes it leads to catastrophic results. Sometime back, WhatsApp announced that it was deleting 1.5 million accounts every month to prevent the spread of fake news.
It has gone hand-in-hand with the rise of the data-driven era – not a coincidence when you consider the…Fake news is a major concern in our society right now. Fake News is any news that is either factually wrong, misrepresents the facts, and that spreads virally (or maybe to a targeted audience). We believe what we saw and sometimes it leads to catastrophic results. Sometime back, WhatsApp announced that it was deleting 1.5 million accounts every month to prevent the spread of fake news.
It has gone hand-in-hand with the rise of the data-driven era – not a coincidence when you consider the sheer volume of data we are generating every second. To deal with fake news detection, a detection system can be made which helps in classifying news as fake or real. For this, Recurrent Neural network can be used as sequence of words matter while differentiating a news as real or fake. The accuracy can be further increased by using encoders - decoders and further by Attention Mechanism. -
Tensorflow vs Keras vs PyTorch vs Theano
Harpreet Singh Sachdev
See publicationThe article discusses the growing popularity of artificial intelligence (AI) since 2016, with a significant number of big companies incorporating AI into their businesses, as highlighted by a McKinsey report from 2018. The potential value of AI is explored across various industries, with estimates reaching $300 billion in banking and $600 billion in retail. The focus then shifts to deep learning frameworks for building AI models, comparing TensorFlow, Keras, PyTorch, and Theano. TensorFlow…
The article discusses the growing popularity of artificial intelligence (AI) since 2016, with a significant number of big companies incorporating AI into their businesses, as highlighted by a McKinsey report from 2018. The potential value of AI is explored across various industries, with estimates reaching $300 billion in banking and $600 billion in retail. The focus then shifts to deep learning frameworks for building AI models, comparing TensorFlow, Keras, PyTorch, and Theano. TensorFlow, known for its dataflow programming and graph computation, is emphasized as the most widely used library, offering scalability and accessibility. Keras is presented as a high-level API for neural networks, capable of running on top of TensorFlow, CNTK, or Theano. PyTorch, developed by Facebook, is praised for its dynamic computational graph and GPU support, while Theano, developed by the Université de Montréal, is lauded for its efficient symbolic differentiation and multi-dimensional array computations. The article concludes by highlighting TensorFlow's prevalence in the industry due to its accessibility, visualization capabilities, and scalability, asserting its popularity in building models, particularly in companies.
-
How does Google Search Mechanism work?
Harpreet Singh Sachdev
See publicationThis article provides insights into the intricacies of Google's search process, explaining the remarkable speed and efficiency with which it delivers results. The author details the steps involved, from accepting and parsing the query to utilizing an extensive index of the internet rather than searching the web directly. The analogy of searching contacts in a phone book is used to illustrate the efficiency of Google's indexing system. The article also highlights Google's vast data centers, each…
This article provides insights into the intricacies of Google's search process, explaining the remarkable speed and efficiency with which it delivers results. The author details the steps involved, from accepting and parsing the query to utilizing an extensive index of the internet rather than searching the web directly. The analogy of searching contacts in a phone book is used to illustrate the efficiency of Google's indexing system. The article also highlights Google's vast data centers, each equipped with numerous servers and automated crawlers, contributing to its impressive lookup capabilities. Additionally, it delves into Google's unique PageRank algorithm, which considers factors like keyword frequency, a page's age, and the number of incoming links to determine search result rankings. Overall, the article demystifies the complex mechanisms behind Google's ability to return 70 million search results in a fraction of a second.
-
How Data Science can be helpful in identifying and curing Mental Illness?
Harpreet Singh Sachdev
See publicationThe article highlights the escalating global mental health crisis, with over 16% of the population affected by conditions such as anxiety and depression. Despite the severity of the issue, more than half of these cases remain untreated, leading to profound consequences for individuals, families, and economies. The shortage of mental health specialists exacerbates the problem, accounting for over 16% of GDP in mental healthcare. The article proposes leveraging Data Science to address the crisis…
The article highlights the escalating global mental health crisis, with over 16% of the population affected by conditions such as anxiety and depression. Despite the severity of the issue, more than half of these cases remain untreated, leading to profound consequences for individuals, families, and economies. The shortage of mental health specialists exacerbates the problem, accounting for over 16% of GDP in mental healthcare. The article proposes leveraging Data Science to address the crisis through innovative methods. These include developing behavioral detection systems using sentiment analysis on social media, AI-powered chatbots for psychometric tests and counseling, parental control apps for early monitoring, and facial expression monitoring systems. The main challenge lies in data collection, with a focus on analyzing internet activities, search history, and content consumption patterns, particularly in the age group of teens to adults up to 40. The integration of AI-powered solutions and data-driven insights is suggested as a means to detect and treat mental health issues in their early stages, potentially revolutionizing mental healthcare.
Projects
-
CI and CD for Machine Learning
-
Data science projects, especially those involving machine learning models, are iterative by nature. Data scientists frequently experiment with different datasets, algorithms, and hyperparameters to optimize model performance. Manual testing throughout this process can be time-consuming and error-prone.
-
Text Autosummarization
-
Build a Text Auto-summarization which can summarize a large text into minimum meaningful sentences keeping the text emotion intact.
The model is built using TF - IDF and Natural Language Processing through NLTK concepts and heapq library of python -
Personal Voice Asssitant
-
Building an intelligent personal assistant (IPA) that can perform tasks or services as instructed through voice commands that can be provided using an python based interface.
Working towards increasing its scope using LSTM model under Recurrent Neural Networks. -
Visiting Card Boundary Detection and Details Extraction
-
Built an algorithm that can open camera and capture a pic. It will then detect for a credit card boundary and finally extract text from it.
This algorithm is built using Contour detection in Opencv and pytesseract for text detection. -
CV Extraction
-
Created an Automated system which can extract every minute details such as Name, Phone No, Email, Location, Educational Qualifications, Gender, Date of Birth, Projects and Skills Set from Resume/Curriculum Vitae using Natural Language Processing and Python when any resume file is fed as input. The code can also be deployed to cloud.
-
Deep Neural Network for Cat Images classification
-
Build a Cat vs Non-Cat image classification by training a deep neural network on it and achieved 91% accuracy. Used basic approach of neural networks and also deployed model using tensorflow.
-
MARKET MIX ANALYSIS
-
• Applying regression techniques to predict Sales of the Company and to set the Sales Target for the future.
• Tools & Technology Used- Python, Regression and Matplotlib for visualization.
-
STEGANOGRAPHY USING PYTHON
-
• Used python programming language to encode and decode secret message into images (hiding text into images and retrieving it without compromising with quality of images).
• Technology Used- Python
-
STACK OVERFLOW SURVEY
-
• Applying Matplotlib and Seaborn to analyze data by cleaning it and visualizing it using several plots.
• Technology Used- Python, Matplotlib and Seaborn
-
SCRAPY AND MYSQL
-
• Using Web Scrapping to collect data from online site and store that data into MySQL tables for further processing.
• Technology Used- Web Scrapping and MySQL
-
FLIGHT CRASH ANALYSIS
-
• Applied Big Data and Hadoop to study the Flight Data recorder data and derive the reason of crash.
• Technology Used- Big Data and Hadoop and PowerBi for visualizing.
Honors & Awards
-
Gold Medal in Graduation
Vivekananda Global University
Won Gold Medal for scoring 9.99 CGPA in my B.Tech
Recommendations received
4 people have recommended Harpreet Singh
Join now to viewWebsites
- Personal Website
-
https://topmate.io/hss245/
- Personal Website
-
https://github.com/hss245
- Company Website
-
https://www.sachdevaisolutions.com/
Other similar profiles
Explore more posts
Explore collaborative articles
We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.
Explore More