jump to: status · repos · research · benchmarks · architecture · activity

I build small, verifiable tools for the parts of the AI stack that fail quietly — agent sandboxes, MCP gateways, model-fit predictors, and post-quantum crypto. Each of the flagship repos below ships with a real test suite, CI, and — where it makes a claim about speed or correctness — a committed benchmark you can re-run yourself.

That's the current chapter. The arc behind it runs from C data structures in 2023, through computer-vision research and a stretch of hackathons and Web3 builds, into GenAI platforms, and now into infrastructure and security. The whole path is below, not just the recent sprint.

• 💼 Now: CTO @ a stealth AI startup (clinical AI) · AI Engineer @ Nextent Labs — groundwater & environmental intelligence for government water departments
• 🛠️ Recently: AI/ML Developer @ RnR Consulting (Delhi) — shipped a model-routing harness that cut inference cost 58% / 65%; Go microservices in a 29-service, Temporal-orchestrated backend serving 500+ concurrent users
• 🎓 Research: 1 Springer Q1 journal · 3 IEEE / SCOPUS papers, plus 2 book chapters in progress — GenAI/GraphRAG, deepfake detection, federated learning
• 🧭 Focus: agent security · AI infrastructure · post-quantum readiness · backend systems
• 🧪 How I work: reference-validated implementations, adversarial tests, reproducible benchmarks — no unverified claims
• 🔜 Shipping next: first-class Node/Ruby bindings for promptproof on its already-public C ABI (the Python ctypes wrapper shipped first); LangSmith/Braintrust export adapters for trace2eval — tracked in each repo's own Roadmap section, not just here
• 🎓 B.Tech (Hons.) CS, Data Science — Symbiosis Institute of Technology, Nagpur (2022–26), CGPA 8.5/10
• 📫 Reach me: LinkedIn · Portfolio · Email

The board above is regenerated daily from the live GitHub API — CI dots, uptime bars, versions and "last commit" ages are real. When a flagship's build is red, it says so.

Small, single-purpose CLIs. Each one exists because a specific check was missing, not because a category needed filling.

The flagships above are the current chapter. This is how it got there. Each row is a real repo, and the honest caveats each one carries in its own README are carried through here too.

Computer vision, accessibility, Web3, and a lot of shipping.

Everything in the flagship tables above, plus:

A unified multimodal GenAI platform integrating GraphRAG multi-agent systems and custom language models for intelligent document processing and knowledge synthesis — Scientific Reports (Nature) 16:16508, April 2026. doi:10.1038/s41598-026-47145-x

Plus 3 SCOPUS-indexed IEEE papers (ICISCT '24 @ Kookmin University, ICPCT '25 @ Amity) and 2 book chapters in progress (federated learning; renewable-energy AI). The deepfake detector above fed one of the IEEE papers; GenAI-Platform is the implementation behind the Scientific Reports work.

# mcp-gateway-lite — reverse-proxy overhead (Apple M4, go1.26.5)
go test -run '^$' -bench . -benchtime=2s ./gateway/...      # 29.3µs direct vs 69.9µs through gateway

# toolcage — per-call WASM sandbox overhead (ubuntu-latest CI, 200 echo calls)
python3 ci/bench.py WORK ./target/release/toolcage x.wasm 200   # 0.415ms median vs 0.089ms unsandboxed floor

# localmodel-fit — prefill 1/params scaling (Apple M4, real ollama)
go run ./bench -model qwen2.5:0.5b -hw m4 -params 494032768     # measured 0.5b/1.5b ratio 3.06–3.18 vs exact 3.125

# ml-kem-rb — hybrid X25519+ML-KEM-768, TLS draft wire format (FIPS 203)
ruby -rml_kem/hybrid -e 'p MLKem::Hybrid.client_init[0].bytesize'  # => 1216 (server share 1120, shared secret 64 B)

Two of my repos (ml-kem-rb, pqc-scan) exist because the crypto deadlines below are real and close. The clock counts down live against these US federal dates (NIST / NSA primary sources).

Go, Rust, TypeScript, Python and Ruby across the flagships — plus C, R, Java and Solidity further back in the arc. The language follows the problem: Rust where a sandbox boundary has to hold, Go for single-binary tooling, Ruby where the ecosystem gap was (a pure-Ruby ML-KEM did not exist), Python for anything touching data.

These aren't scattered side-projects. The Ideal Harness is the control plane the whole path runs through; modelgate gates the LLM calls; mcp-gateway-lite filters the tool calls; toolcage sandboxes each one; mcp-sentinel, agent-rules-audit, agent-flightbox and trace2eval watch the run — and ml-kem-rb / pqc-scan are the post-quantum floor the whole thing has to stand on.

Node size is real commit count. The MCP hub links the repos that actually speak the protocol; the bright edge is the CI cross-check where mcp-gateway-lite verifies mcp-sentinel's own lockfile output — a real test, not a diagram flourish.

── EOF ────────────────────────────────────────────────────────────
   if a tool makes a claim, it ships with the test that proves it.
────────────────────────────────────────────────────────────────────