Python-native · B.Tech CSE (Cybersecurity), IIIT Kottayam, batch of 2028
I build and measure security tooling for AI agents, and I build the agent and ML systems themselves.
The tooling is written in Python. Across my public repositories, the code that is not Python is limited to front ends (the Next.js dashboard in soc-parallax and a few single-file HTML demos) and one Rust file used as a test fixture in pqc-scanner.
My security background (detection engineering, application security, threat modelling) shapes how I build AI systems: treat model output and tool metadata as untrusted input, enforce controls in code rather than in the prompt, and publish what a detector misses next to what it catches.
Open to: internships May-Jul 2027, full-time roles from 2028 (AI security, AI engineering, application security; India or remote).
| Project | What it is | Measured result |
|---|---|---|
aisec-suite (PyPI 0.11.0) |
CI security gate for what AI agents trust: MCP tool metadata, agent memory files, RAG documents. SARIF output, baselines, reachability analysis, an MCP runtime guard and an LLM red-team CLI. | Structural poisoning detector: 53.7% recall on my held-out attacks, 14.0% on other researchers' attacks (MCPTox, MCP-Poison-Bench), 0.05% false positives on 3,974 real tool descriptions (2 flagged). It misses roughly half of unseen attacks, and the README says so. |
| mcp-scan-study | Hand-labelled study of 135 public MCP-server repos, with the scanner patches that came out of it. | No real tool poisoning found in the wild. mcpaudit precision went from 37.5% to 64.3% on a held-out sample, and from 40% to 72.7% on a blind-labelled fresh corpus. |
| mcpaudit · memsentry · ragsentry | The three scanners behind aisec-suite: MCP tool surfaces, agent memory/context files, RAG source documents. | On attacks written by others, memsentry and ragsentry catch the stock "ignore previous instructions" phrase and 0-3% of attacks without it. |
| soc-parallax | Detection platform: per-entity behavioural baselines, an attributable risk score, MITRE ATT&CK mapping from a versioned rulebook, incident correlation in Neo4j, grounded narratives. FastAPI, PostgreSQL, Next.js. | On 278 real EVTX attack samples, 29.5% reach medium severity or higher (27.3% on a held-out half). Nothing reaches high on held-out benign data. An earlier inflated 53% figure was found in an audit and replaced. |
| osiris | Model-agnostic assistant runtime: 8 LLM providers behind one failover router, tool-scoped agents, embedding memory. I use it daily. | Traced a local-model slowdown to Ollama's thinking mode and fixed it with a native adapter (about 100 s to 10 s per call, measured live). |
More in the portfolio: PhantomGrid (behavioural biometrics for banking fraud, team hackathon project), promptstrike (multi-turn LLM red-teaming), loopcheck (agent-loop waste auditor), darkdecoder (ATT&CK + ATLAS threat-intel analysis).
7 pull requests merged and 13 open at the last check (13 Nov 2026). Live status is on each PR; the full table is on the portfolio.
| Merged | What it fixed |
|---|---|
| tobymao/sqlglot #8780 | Parser._replace_lambda shared one DataType node between several casts. |
| PrefectHQ/fastmcp #6063 | Host callbacks were not counted against the Monty sandbox time limit; added a wall-clock timeout and a call budget. |
| BerriAI/litellm #47091 | A string stop value is now wrapped for Cohere stop_sequences. |
| sqlfluff #8899, #8901, #8902, #9059 | Dialect grammar fixes for StarRocks, Doris and MySQL. |
Open work includes authorization fixes in LiteLLM (#47105), MCP pagination in CrewAI (#8240), and fixes in PennyLane, Qualtran, Graphene and Presidio.
- Language: Python (asyncio, FastAPI, Pydantic, pytest, AST analysis, scikit-learn)
- AI engineering: LLM tool-calling and agent loops, provider routing and failover, embeddings and retrieval, evaluation harnesses, local models with Ollama
- AI security: MCP security, prompt injection and tool poisoning, RAG and memory poisoning, LLM red-teaming (PAIR, TAP, Crescendo), OWASP Top 10 for LLMs, runtime guards, SARIF
- Security foundations: MITRE ATT&CK and ATLAS, Sigma, threat modelling (STRIDE), detection engineering, Z3 model checking
- Data and infra: PostgreSQL, Neo4j, SQLite, Docker, GitHub Actions
- Practice: TryHackMe top 2% globally
- Static scanners miss most MCP tool poisoning (Nov 2026): what my detector catches, what it misses, and why a learned classifier did not close the gap.
Every number links back to a script, a labelled dataset or a study document. Where a result is weak, the README says so: single labeller, synthetic data, held-out sets evaluated more than once. AI coding assistants were used on some projects and are disclosed in those READMEs.


