I build agentic, retrieval, and evaluation systems, plus the applied-ML and full-stack products around them. 4+ years across production data, ML, and applied AI.
| Project | Stack | What it does |
|---|---|---|
| Autonomous Impact Analyst | LangGraph · Claude · Neo4j · sqlglot · dbt | A LangGraph agent that traces a warehouse change through a 451-node column-level lineage graph, scores blast-radius risk with a deterministic formula, and opens a sqlglot-validated fix PR. The LLM only writes the summary. |
| SEC RAG | FastAPI · Qdrant · BM25 · XBRL · Claude · MCP | Grounded SEC-filings answer engine that enforces span citations in code, refuses on weak retrieval, pulls figures from structured XBRL, and gates releases on a golden set. 106 tests, deps faked behind protocols. |
| LedgerBench | DuckDB · sqlglot · dbt · Anthropic · OpenAI | Open-source benchmark showing analytics agents run SQL 100% cleanly yet are business-correct on only 9–59% of answers. Every number traced to a committed run manifest. |
| sliceval | scikit-learn · PyPI | Published pip install sliceval library that finds underperforming data subgroups with bootstrap confidence intervals and permutation significance tests. 162 tests across 13 model types. |
| Neo4j GraphRAG → Apache Hamilton | Apache Hamilton · Neo4j | A 4-module GraphRAG pipeline (ingest → embed → retrieve → generate) contributed and merged upstream as apache/hamilton PR #1532. |
| Cover Drive | PyTorch · QLoRA · Unsloth · FastAPI · Next.js | A fine-tuned Qwen2.5-1.5B (QLoRA) engineered so it cannot state a wrong fact: deterministic fact-fill plus validate-or-fallback serving. 190 tests. |
🚧 Now building — MAIRS: a multi-agent investment-research layer over the SEC grounding engine.
More work, and an "Ask the Projectionist" AI guide, at kmandhar.com.
| Languages & Compute |
|
| ML & LLM |
|
| Data & Databases |
|
| Cloud & Serving |
|
AWS Machine Learning – Specialty · Azure Data Engineer Associate (DP-203) · Azure AI Engineer Associate