01AI Work

Built for the real world.

Production AI at Optum and InfoCepts, and CaseCraft — a language model fine-tuned for India's Insolvency and Bankruptcy Code. Client identities are withheld.

2 enterprise chapters1 public research modelClient identities withheld

01Enterprise work

Production AI, in two chapters.

Jun 2025 — Present

Optum · UnitedHealth Group

Software Engineering Lead, AI/ML

01
  • Led an AI-orchestrated automation platform: six FastAPI services on Google Kubernetes Engine, with Vertex AI document extraction.
  • Built for people in the loop — a React review interface and BigQuery audit and staging, with human approval before any authoritative write.
  • Keyless CI/CD across four environments, gated on at least 90% test coverage.
FastAPIGKEVertex AIBigQueryReact
Engineering deep dive
Apr 2022 — Jun 2025

InfoCepts

Senior Data Scientist · Lead, Generative AI COE

02
  • Production ad-inventory forecasting with Prophet and PySpark on Amazon EMR, over 2TB+ of data across seven brand verticals.
  • Conversational BI, enterprise search and multimodal PDF retrieval with question answering.
  • Led the Generative AI centre of excellence; SME and mentor for the data-science bootcamp.
ProphetPySparkAmazon EMRRAGMultimodal

02Public research · M.Tech dissertation

CaseCraft

A language model adapted to India's Insolvency and Bankruptcy Code.

The problem

India's Insolvency and Bankruptcy Code is a dense, evolving body of statute, regulation and case law. General-purpose language models are not tuned to its vocabulary, structure or the way its documents refer to one another.

My contribution

My M.Tech dissertation at BITS Pilani (AI & ML, NLP specialisation, 2023–2025). I fine-tuned the model and published it on Hugging Face. The source repository holds the document-collection scripts, fine-tuning code and a RAG pipeline.

Technical choices

Incremental fine-tuning with LoRA on Apple MLX, starting from the Llama 3.2 11B Vision Instruct base in 8-bit. Training ran locally on an Apple M3 Max with gradient checkpointing and adaptive chunking.

Base model
Llama 3.2 11B Vision Instruct · 8-bit (MLX community build)
Method
Incremental LoRA fine-tuning on Apple MLX
Corpus
1,200+ legal documents — IBC judgments, amendments and laws
Hardware
Apple M3 Max · gradient checkpointing · adaptive chunking

Results

Evaluation results are not yet published. Until they are, this page makes no accuracy or performance claims.

Limitations

Scoped to Insolvency and Bankruptcy Code material. It is a research artifact, not legal advice and not a substitute for counsel.

Public artifacts

Have a problem worth thinking through?

AI systems, research ideas and thoughtful products.

Let's talk