NYC · RemoteBlackRock AI Accelerator

Garvit Khurana

AI Product Lead & Engineer

Shipping LLM and RAG systems on live capital-markets workflows — model, infra, and finance-grade evals in one loop.

Now

In the model layer of markets.

Bond ops used to retype terms out of 100+ page prospectuses by hand — slow, error-prone, and the bottleneck on getting a bond to market.I own the RAG workflow that turns those prospectuses into structured security records, cutting time-to-market ~40%, then extended the same loop into private credit where the docs are messier and naive extraction breaks.

Day to day: roadmap and delivery across model (RAG, prompting, extraction), infra (APIs, pipelines, Azure), and evals/tracing so bad fields fail loudly before they reach a client workflow.

Model

OpenAI / Claude · LangChain · RAG · ChromaDB · Structured outputs · Tool calling

Infra

Python · FastAPI · SQL · PySpark · Azure / AWS · Databricks · ETL

Evals / Tracing

OpenTelemetry · Azure Monitor · LangSmith · Golden-set field checks

Builds

Outside the day job.

Building now

Altline

The model reads. The engine verifies.

LLM extracts terms from a credit agreement; a rules engine recomputes borrowing-notice and rate-set numbers instead of trusting what an administrative agent sends. Agentic back office for alternative assets.

LLM extraction · Guardrails · FastAPI

Public

agent-relay

Hands an in-progress coding task to the next available agent (Claude, Codex, Cursor) when one hits a rate limit — git worktrees carry the file state so you don’t restart the prompt from scratch.

Git · CLI · Multi-agent

Public

mlx-voice-clone

Local voice-clone TTS on Mac (Qwen3 / MLX) with a FastAPI UI — a personal project on on-device inference.

MLX · Qwen3 · FastAPI

Work

Where it’s shipped.

2024 — Now

Product Lead / Engineer II, AI Accelerator · BlackRock

  • Owned 0→1 for an internal LLM/RAG workflow that turns long bond prospectuses into structured security records — less manual retyping, ~40% faster time-to-market for new public bonds.
  • Extended into private credit: tightened model → infra → evals/tracing so field validation and traced extraction runs fail loudly before bad data hits a client workflow.
  • Work with client and RM teams on adoption; partner with data science on SQL, telemetry, and Azure so prioritization tracks what breaks in use.

2022 — 2024

Private Markets Associate · BlackRock

  • Launched the Private Governance Platform after teams were catching data errors late and separately: −80% manual intervention, +20% reporting accuracy into Aladdin.
  • Built NLP issuer-search so PMs weren’t hunting the same issuer across systems that didn’t talk to each other.

Earlier

Product · data · ML · APG · PadSquad · HPE

  • ESG + NLP on 13F filings at APG; growth product in AdTech; CNN assistive-tech at HPE.
  • Columbia MSBA · VIT CSE · NUS Data Science · four ML research publications.

Contact

If you’re building with AI.

Open to roles where the product is AI in a complex, high-stakes domain — NYC, California, US remote, London, Canada.

garvit.khurana@columbia.eduResumeLinkedInGitHub