Currently: 7th semester @ PES University, Bengaluru | Expected graduation 2027 | Open to AI/ML engineering internships

Most of what I build sits where retrieval, graph structure and inference meet, because those are the pieces that break in production and need thinking through.

Temporal knowledge graph over a codebase: every node is a file, function or symbol, and every edge carries the timestamp of the commit or PR that introduced it. Questions like "why was this module rewritten" are answered from real history, not guessed. Retrieval runs TKG traversal, BM25 and vector search in parallel, merged by a cross-encoder reranker.

• Insight: Commits and PRs are inherently temporal; a TKG plus a general LLM beats specialised code models for history questions
• Tech: Python, FastAPI, Kuzu, ChromaDB, BM25, tree-sitter, Groq, LLaMA 3.3-70B, React, Vite, Tailwind

Give it a migration and it tells you which queries break, which silently return wrong results, what locks and for how long, and how to roll back. An eight-node LangGraph agent with per-node LLM retry and a deterministic fallback, so the analysis completes even with no LLM configured. Risk comes from a gradient-boosting model (SHAP-explained), a Dirichlet severity posterior that updates from recorded outcomes, and an LLM verdict. 88 tests.

• Insight: The LLM adds a verdict, it never gates the result. If it is down you still get the report
• Tech: Python, FastAPI, PostgreSQL, LangGraph, scikit-learn, SHAP, PyMC, pglast, Docker, React

Hourly deterioration prediction on real ICU stays. I own evaluation: a 500-case, 14,000-call harness that exposed a consistency illusion in naive multi-agent debate and validated a gating fix (decision-alignment fidelity 0.70 → 0.89), the cost-aware BigQuery extraction layer, and the pytest suite that enforces patient-level splits and leakage-safe feature windows. Team ensemble: 0.788 AUROC on the held-out test set.

• Tech: Python, BigQuery, PyTorch, scikit-learn, pytest

Isolation Forest over CNC sensor data, with each anomaly batch handed to an LLM for a plain-English diagnosis.

• Tech: Python, FastAPI, scikit-learn, Groq, LLaMA 3.3-70B, Streamlit

Raft in Go: leader election, WAL-fsynced log replication, and single-RPC catch-up across a 3-replica cluster.

• Tech: Go, gRPC, Protobuf, Docker

Languages: Python · Go · SQL · TypeScript/JavaScript · C LLM & ML: RAG · LangGraph · LangChain · MCP · PyTorch · scikit-learn · SHAP · PyMC · sentence-transformers · Groq · LLaMA · HuggingFace Backend: FastAPI · gRPC · Protobuf · REST · WebSockets · Docker Data: PostgreSQL · BigQuery · SQLite · Kuzu · ChromaDB · Redis Tools: Git · pytest · Linux · k6 · Prometheus

• Portfolio: shrishthakur.vercel.app
• LinkedIn: linkedin.com/in/shrish-singh-thakur
• Email: shrishsinghthakur04@gmail.com
• Location: Bengaluru, Karnataka, India