Technical Lead with 9 years of experience — LangGraph, Azure OpenAI, RAG pipelines, and cloud architecture at enterprise scale.
Architected a federated RAG platform where each tenant maintains an isolated private knowledge base while optionally subscribing to a shared global knowledge layer (UMBRU). A single query dynamically retrieves from both sources, merges results, and delivers a unified, context-aware response — enabling knowledge sharing at scale without data boundary violations.
Built a fault-tolerant, multi-agent SDLC orchestrator that turns an FRS into a validated, machine-readable ProjectBlueprint through 5 specialist agents and 2 human-in-the-loop checkpoints. Solves the "chat-log output" problem of typical agentic dev tools: every HITL decision is timestamped, linked to the specific feature or edge case it resolved, and every agent output is Pydantic-validated before the state machine allows a transition — so a bad LLM output triggers a retry, never a silent pass-through. LangGraph checkpoints every node, so a crashed process resumes exactly where it stopped with zero duplicated work. The differentiator over "just chaining LLM calls": the state machine, not the model, owns sequencing and enforcement.
Built an agentic AI governance and compliance platform that evaluates AI systems against major regulatory frameworks (EU AI Act, ISO 42001, NIST AI RMF) through automated control checks and evidence generation, rather than manual compliance spreadsheets. Aimed at teams shipping AI systems who need a repeatable, auditable way to demonstrate governance coverage instead of ad-hoc documentation per audit cycle. [DRAFT: exact automation scope, control-mapping method, and end-user workflow not verified against source — reconstructed from a one-line project summary only.]
Built an open-source evaluation and red-team harness purpose-built for LangGraph agents, addressing the gap where most agent frameworks ship without systematic tooling to adversarially test or behaviorally evaluate the agents built on them. v1.0.0 covers structured evaluation runs; a V2 in development adds a provider-agnostic package for testing real (non-mocked) LLM-backed agents rather than only simulated ones, closing the gap between "passes in the eval harness" and "behaves correctly against a live model." [DRAFT: exact evaluation methodology, scoring approach, and adversarial test categories not verified against source.]
Built a platform for systematically comparing RAG configuration choices — retrieval strategies, chunking approaches, embedding models — instead of tuning a RAG pipeline by trial and error against production traffic. Reached v1.0.0 across 7 tagged releases with 2,082 passing tests, indicating a mature, heavily-tested configuration/experimentation surface rather than a single fixed pipeline. [DRAFT: exact experiment-tracking mechanism, supported retrieval strategies, and comparison/reporting UI not verified against source.]
Built a production-grade Graph RAG system that automatically ingests papers from arXiv, extracts entities and relationships, and exposes the resulting knowledge graph through an interactive D3.js visualization — letting a user explore paper-to-paper connections visually rather than only through text search. Aimed at researchers or technical readers who want to navigate a research area by its actual citation/concept graph instead of a flat list of retrieved chunks. [DRAFT: exact entity/relationship extraction method and graph storage layer not verified against source.]
Built a deterministic criterion-scoring pipeline for Indian equity market analysis, architected to keep factual computation (a deterministic layer) strictly separate from natural-language explanation (an LLM narration layer) — so the analysis numbers are never something an LLM could hallucinate, only narrate. This mirrors a broader architectural principle used across other projects: the model proposes/explains, a deterministic layer decides/computes. Status: in progress. [DRAFT: exact scoring criteria, data sources, and narration prompt design not verified against source.]
Built a Python library that reconstructs tables spanning multiple PDF pages into a single unified structured dataset — addressing a common, unsolved gap in standard PDF-parsing libraries, which typically treat each page's table fragment as independent and require manual stitching. 132 tests at 99% coverage indicate a mature, well-tested parsing core rather than a quick script. [DRAFT: exact table-boundary detection heuristic and supported PDF table formats not verified against source.]