AI/ML Engineer — LLM evaluation infrastructure & agentic systems
- AI/ML Engineer Intern at CobuildX.ai: agentic workflows with LangGraph and MCP, and a major contributor to the ai01 LLM evaluation platform.
- Built evaluation tooling: async eval pipelines, multi-agent trajectory evaluation, Hungarian-matched entity-extraction benchmarking, dataset versioning and run comparison.
- Former Graduate Technical Intern (AI Engineering) at Intel: multi-agent LLM automation and GraphRAG.
- Open-source contributor to Meta FAISS (merged) and Google DeepMind Gemma (open PR).
Portfolio · LinkedIn · Email · ai01-eval on PyPI
| Project | What it is | Highlights |
|---|---|---|
| extract-eval | Structured-extraction evaluation library | Hungarian line-item matching, bootstrap CIs, run comparison, CI gating. CORD-v2 benchmark of Donut: micro-F1 0.876 (95% CI 0.842–0.908) on 100 docs |
| rewind (live) | Time-travel debugging for AI agents | Record, fork and replay runs; trajectory evaluation; MCP server |
| autonomous-platform-intelligent-agent | Natural language to GitHub operations agent | Learns API constraints; hardened code-synthesis sandbox; 64 tests |
| PoliRAG | RAG for policy compliance | Hybrid BM25 + FAISS retrieval with cross-encoder reranking; Streamlit demo |
| aws-ml-risk-scoring-service (live) | XGBoost credit-default scoring API on AWS | AUC 0.77 |
| WaferDefectX | Wafer defect detection (computer vision) | INT8 quantization 629 KB to 237 KB; ONNX / OpenVINO |
- AI/ML Engineer Intern, CobuildX.ai (Jun 2026 – present): LangGraph + MCP agent workflows; ai01 evaluation platform (AWS Lambda + SQS async pipeline, multi-agent evals, BYOK via litellm, published SDK).
- Graduate Technical Intern, AI Engineering, Intel (Jul 2025 – Feb 2026): multi-agent ticket automation on Azure AI Foundry + ServiceNow (input tokens down 39.4%, model cost down 35–45%); GraphRAG with LlamaIndex, Neo4j and FAISS (retrieval accuracy up 32%, lookup latency down 45%).
- Meta FAISS — PR #4792, merged (CI build fix).
- Google DeepMind Gemma — PR #590, open PR (Gemma2 sampler crash fix).
- Patents — 3 invention disclosures (IDF-B) prepared with the VIT IPR & TT Cell: multi-agent GenAI for legal contract drafting; SAR semantic downlink with onboard VLMs; hallucination-hardened mission-briefing engine. Status: disclosures only, no patent filed or granted.
Python, PyTorch, LangGraph, LangChain, LlamaIndex, MCP, FAISS, Neo4j, FastAPI, AWS (Lambda, SQS, DynamoDB), Docker, Kubernetes, React/TypeScript.
M.Tech CSE (AI & ML), VIT Vellore, 2024–26 (CGPA 9.34) · B.E. CSE, Loyola-ICAM, 2020–24 (CGPA 9.16)
- Winner, TECHNO TEEN Science Fair (DigiPass)
- Top 100, Niral Thiruvizha, Government of Tamil Nadu (AgroVoiceAI)
- Finalist, Ashok Leyland Hackathon