I design and ship production GenAI systems — the messy, end-to-end kind, not just demos. MSc Artificial Intelligence (Distinction) from Queen Mary University of London, 4+ years of engineering experience, and a track record of taking ambiguous problems through architecture, deployment, and adoption.

Most recently I co-founded Scintilink (Dec 2024 – Sept 2025), a multi-agent AI platform that served 100+ researchers in production. Before that, I spent three years at Lumen Technologies shipping AI-driven forecasting tools that informed $25M+ in annual network investment decisions.

I care about AI that's safe, reliable, and actually useful — not just impressive in a screenshot.

An 8-agent system orchestrated with LangGraph, deployed on GCP, helping researchers reason over scientific literature with citation-backed answers.

Fine-tuned LLM + Multi-Agent RAG + Graph Neural Networks

• Fine-tuned Microsoft Phi-4 (14.7B params) with LoRA on FinQA → +550% accuracy on financial QA
• 3-way LangGraph query router over a 68,989-embedding Qdrant pipeline
• Streaming ingestion of 18,000+ SEC EDGAR filings → knowledge graph of 17,552 nodes / 420,796 edges
• Four GNN architectures: GraphSAGE, GATv2, Temporal GNN, Global Graph RAG

Cost/latency-aware model routing with confidence escalation

A LangGraph pipeline that classifies intent, scores mission-criticality and latency-sensitivity, then picks the optimal model and deployment target — with 0.65 confidence-threshold escalation, intent-aware fallbacks, and full Pydantic pipeline traces for observability.

Hybrid retrieval over BMW, Tesla, and Ford annual reports

BM25 (30%) + semantic (70%) fused via Reciprocal Rank Fusion across 2,871 chunks → 85.9% keyword match, 100% pass rate on 13-case eval suite.

Sub-500ms search on 100k+ documents · BM25Okapi · multi-threaded.

Agentic & LLM — LangGraph · LangChain · LangSmith · HuggingFace Transformers · LoRA · PEFT · OpenAI · Anthropic · Groq · FlashAttention-2

RAG & Retrieval — Hybrid retrieval (BM25 + semantic + RRF) · Qdrant · ChromaDB · Pinecone · Multilingual-E5-Large · knowledge graphs

Evaluation & Observability — Custom eval harnesses · A/B testing · agent decision logs · telemetry dashboards · regression testing

mindmap
  root((What I'm digging into))
    Agentic AI
      Multi-agent coordination patterns
      Tool-use planning and verification
      Long-horizon task decomposition
    Evaluation
      Agent decision logging
      Failure-case taxonomy
      LLM-as-judge calibration
    Retrieval
      Hybrid search tuning
      Query decomposition
      Contextual compression
    Safety and Reliability
      Guardrails for production
      Human-in-the-loop design
      Regression testing for LLMs

MSc Artificial Intelligence — Queen Mary University of London (2024–2025) · Distinction
B.Tech Computer Science & Systems Engineering — KIIT University (2017–2021)

Open to: Applied AI / AI engineer roles focused on agentic systems, RAG, and production LLM work. Based in London, UK — happy to discuss remote, hybrid, or relocation.

• 💼 LinkedIn — neelay-choudhury-768537152
• 📧 neelaychoudhury1999@gmail.com