Agentic AI security · Cloud-native systems · Full-stack
I run empirical safety research on autonomous agents and ship the tooling that operationalizes it. Current work spans seven attack categories (C1-C7), tested across 29 models and 8 providers; the findings feed into production-ready detectors, evals, and red-team plugins contributed to garak, llm-guard, deepeval, promptfoo, and inspect_evals.
- Kogni·OS
: Python agent framework built from scratch. SQLite-native RAG, readable ReAct loop, streaming, async, eval harness, HTTP serve, 9 providers (Anthropic, OpenAI, Groq, Gemini, Mistral, Cohere, Ollama, Bedrock, xAI)
- Bolo: bilingual speech and vocabulary app for toddlers (English + Hindi, Web Speech API, PWA)
- geo-pulse: signal-first geopolitics briefs with automated hourly updates (Python, GitHub Actions)
- BVSA: site for a grassroots nonprofit in Bundelkhand, built as a PWA for slow rural connections
- Planning Poker: real-time agile estimation tool; cards flip simultaneously so nobody anchors first (Firebase Realtime DB)
Languages Node.js · Go · C# · Python · SQL
Infra Kubernetes · Docker · cloud-native · GitHub Actions
AI / Agents Claude Code · MCP · agentic workflows
AI safety / agentic security (Dwivedi 2026, preprint: empirical research across 29 models and 8 providers, C1-C7)
| Repo | PR | What |
|---|---|---|
| NVIDIA/garak | #1918 | Multi-turn persona injection probe + detector (C2) |
| NVIDIA/garak | #1920 | Fix: retag AgentBreaker owasp:llm07/08 to owasp:llm06 |
| NVIDIA/garak | #1921 | Fix: coerce non-string REST response fields to str in RestGenerator |
| NVIDIA/garak | #1925 | C5 multi-agent orchestrator trust exploitation probe |
| promptfoo/promptfoo | #10008 | persona-injection redteam plugin (C2) |
| promptfoo/promptfoo | #10020 | orchestrator-trust-injection redteam plugin (C5) |
| protectai/llm-guard | #350 | AgentEscalation output scanner (C4) |
| protectai/llm-guard | #351 | AgentMemoryPoisoning output scanner (C6) |
| protectai/llm-guard | #352 | CredentialExfiltration output scanner (C7) |
| confident-ai/deepeval | #2855 | AgentEscalationMetric LLM-judge metric (C4) |
| confident-ai/deepeval | #2863 | AgentMemoryPoisonMetric LLM-judge metric (C6) |
| UKGovernmentBEIS/inspect_evals | #1906 | Register: AgentEscalationEval task (C4) |
| UKGovernmentBEIS/inspect_evals | #1907 | Register: OrchestratorTrustExploitationEval task (C5) |
| openclaw/openclaw | #97107 | Fix: bypass fail-closed interpreter heuristic when security=full |
Other
| Repo | Contribution |
|---|---|
| github/awesome-copilot | 2 merged PRs: technical job-search skill, technical interview prep coaching agent |
| santifer/career-ops | 8 merged PRs: git-safety fixes, Windows symlink handling, dashboard scripts, rollback guard, interview-prep contract coverage |
| AppMetrics/AppMetrics | Add IWebHostBuilder extensions for .NET metrics |
| minio/minio-dotnet | Fix stream disposal bug in ToXML method |
