I build AI systems that move from evidence to decisions: governed agents, retrieval systems, graph intelligence, fine-tuned models, and APIs that are tested, observable, and deployable.
My work sits at the intersection of applied AI and dependable software engineering. I care about the hard parts that make systems useful in the real world: provenance, retrieval quality, deterministic validation, stateful orchestration, failure recovery, privacy boundaries, and deployment on constrained hardware.
- Open to AI Engineer opportunities spanning Agentic AI, GraphRAG, retrieval systems, LLM platforms, and efficient model deployment.
- Building production Agentic RAG infrastructure for legal research with LangGraph ReAct workflows, Zilliz hybrid retrieval, and a dedicated FastMCP server that exposes legal intelligence as native tools.
- Exploring QLoRA/DoRA fine-tuning and low-cost inference for capable models on constrained hardware.
Reconstructing six years of food commerce into an auditable personal economic- intelligence graph. Lunarbit turns invoices, order histories, and email orders into a temporal Neo4j knowledge graph, then answers financial questions with hybrid retrieval and citation-level evidence.
- 454 reconstructed orders · 53k+ graph nodes · 85k+ relationships
- Neo4j GraphRAG with exact matching, Lucene/BM25, HNSW vector search, RRF, Cohere embeddings/reranking, and bounded graph traversal
- LangGraph stateful workflows with durable conversation checkpoints, guarded tool use, deterministic Decimal-safe reconciliation, and privacy-safe public projections
- FastAPI contracts, streaming answers, source provenance, typed validation, container smoke tests, CodeQL, and CI-enforced test-driven development
Lunarbit is designed to answer questions such as: Which restaurants account for the most orders? How has a dish price changed over time? How much did fees, discounts, and membership economics change my actual spend? Every answer is constrained by the graph and its evidence rather than generated from plausible text.
A 4.5B multimodal model fine-tuned with QDoRA on 37k+ grounded training
samples across 11 real-patient dataset families. The pipeline separates raw
truth, deterministic structure, and grounded generation to reduce unsupported
clinical claims. Quantized GGUF artifacts run fully offline through
llama.cpp/Ollama on low-cost hardware, with Tamil/Tanglish support and
calibrated confidence outputs.
An evidence-grounded document compiler for truthful, tailored, exactly one-page resumes. TITAN converts a job description and verified career evidence into typed JSON, renders a locked LaTeX template, compiles a PDF with Tectonic, and validates the artifact recruiters and ATS systems actually see.
- Provenance attached to claims, projects, skills, and experience
- Template-aware space planning, ATS reading-order checks, hyperlink checks, geometry measurement, and page-utilization gates
- Bounded, element-level repair loops instead of unconstrained regeneration
- Three reviewed A4 templates, deterministic validators, provider fallback, and a test-first Python architecture ready for LangGraph/HITL expansion
A framework-independent asynchronous ReAct system for end-to-end job discovery. Seven specialized agents coordinate twenty tools through a custom runtime, with multi-provider routing, automatic failover, Qdrant semantic memory, Supabase structured memory, and Telegram delivery.
- Multi-Agent RAG Chatbot — async message-bus orchestration over heterogeneous document ingestion and retrieval.
- Business Document Processing Workflow — specialized-agent document automation with structured outputs.
- Computer Vision Player Tracker — YOLOv5 tracking with Kalman filtering and Hungarian assignment for robust re-identification under occlusion.
- Advanced GenAI Server — a headless Ubuntu AI server with Ollama, LangChain, Nomic embeddings, and MySQL-backed conversational memory.
- Portable Server — dual-boot Linux infrastructure built from a repurposed laptop for wireless, always-on hosting.
Python · LangGraph · GraphRAG · Neo4j · FastAPI · FastMCP ·
Cohere · BM25 · HNSW · RRF · Qdrant · Zilliz/Milvus · PyTorch ·
QLoRA/DoRA · GGUF · llama.cpp · Docker · CI/CD · TDD
- LLMs propose; deterministic systems decide. Money, graph truth, provenance, privacy, and release gates are code-owned.
- Evidence before eloquence. Unsupported claims are rejected or surfaced for review.
- Production behavior starts with tests. Contracts, integration paths, failure modes, and security boundaries are continuously exercised.
- Design for failure. Provider fallback, bounded retries, checkpointing, sanitized errors, and explicit abstention are first-class behavior.
Building systems that can explain where an answer came from—and still work when the model, network, or input is imperfect.