Late City Edition theomthakur.github.io Open to full-time roles · U.S.-wide
Applied AI · full stack · shipped to real users

The Om Thakur Times

"All the systems fit to ship"
Software engineer, reporting from coast to coast
Vol. I · No. 1 New York, NY · Open to Relocation Price: A coffee chat
Résumé ↗ Email Book a chat GitHub LinkedIn
Profile · Now Available for Hire

He Builds the AI, Then Builds the Thing That Proves It Works

Om Thakur builds applied AI and full stack products end to end, from the data layer to the interface, and builds the evaluation that proves they work. Open to full-time roles anywhere in the U.S.

Om Thakur builds software people actually use: retrieval pipelines, agentic workflows, and the full stack products around them. Having completed his Master's in Computer Engineering at NYU (conferred May 2026) and wrapped a year building medical AI at NYU Langone Health, he is looking for his next full-time role.

His work runs from the interface down to the model and back. At NYU Langone he built a clinician-facing retrieval system over 170 clinical documents on a HIPAA-compliant in-house vector store, then built the harness that graded it: ten foundation models across six providers, 1,950 runs scored on latency, cost, faithfulness and citation quality. It caught silent failures in citation grounding that no demo would have surfaced.

"Anyone can demo a model. The work is proving it still holds on the thousandth query."

That belief is now packaged and public. benchtrace on PyPI measures whether one model is actually better than another and refuses to name a winner when the evidence does not support one. agentcassette on npm records an agent run once so a test suite stops paying an API bill to learn the same thing twice.

He takes projects end to end, from a vague idea or a detailed spec through design, testing and production-hardening, solo or on a team, and has worked with clients from first ideation through shifting requirements. The same year at Langone he stood up HPC pipelines moving 50+ terabytes of clinical sequencing data across NVIDIA A100 GPUs.

Earlier, at 4iBiz, he shipped SaaS to more than 20,000 users, cut dashboard latency in half across a thousand gyms, and delivered 18 production features on an eight-engineer team. Underneath all of it sits a systems habit: a message queue written in Rust, a consensus store machine-checked in TLA+. Understanding how things fail is what makes the rest trustworthy. Off the clock, he is found photographing the city on a 35mm, on the cricket pitch, or mentoring students at NYU Tandon.

Status
Available now
Location
Anywhere in the U.S.
Work authorization
F-1 STEM OPT
Sponsorship
Not needed until June 2029
On the Job

Building Medical AI at NYU Langone

Student Research Intern · AI Systems & Cloud Infrastructure

  • Designed an LLM evaluation framework benchmarking 10 models across 6 providers (OpenAI, Anthropic, Bedrock-hosted Llama/Claude) on latency, cost, faithfulness, and citation quality across 1,950 evaluations — surfacing silent failure modes that informed production model selection.
  • Built a clinician-facing RAG pipeline over 170 clinical documents with a HIPAA-compliant in-house NumPy vector store, sentence-aware chunking, and multi-provider inference fallback.
  • Wrote HPC pipelines processing 50+ TB of clinical sequencing data across 4 cohorts on NVIDIA A100 GPUs under SLURM; built a schema-validated ETL into Parquet on S3 with Glue and Athena.

LLM EVAL · AWS BEDROCK · OPENAI · ANTHROPIC · SLURM · FASTAPI · PYTHON

Four Platforms, 20K Users

Software Developer · Pune, India

  • Shipped 4 production SaaS platforms end-to-end with Angular, React, Node.js and MongoDB.
  • Built Fytrack Corporate from scratch — architecture, schema, API layer, frontend.
  • Architected REST APIs with JWT auth, RBAC and validation middleware.
  • Worked directly with founders and clients — from ideation through changing requirements to launch.

ANGULAR · REACT · NODE.JS · MONGODB

From Intern to Full-Stack Owner

Software Developer Intern · Pune, India

  • Contributed to 3 production platforms; grew from UI tasks to owning full-stack features.

REACT · ANGULAR · NODE.JS

Education & Research

Degrees & Publication

  • New York University — MS Computer Engineering · GPA 3.9 · conferred May 2026
  • Vishwakarma Inst. of Technology — B.Tech Computer Engineering · GPA 8.86 · 2019–23
  • Peer-reviewed: "Solar Panel Damage Detection & Localization of Thermal Images"J. Failure Analysis & Prevention, Aug 2023. YOLO-based thermal imaging, 69% mAP. read →

MACHINE LEARNING · DEEP LEARNING · ML SYSTEMS · BIG DATA · COMPUTER VISION · SYSTEMS ARCHITECTURE

The Toolkit

What He Travels With

AI & LLMs — RAG, agentic workflows, evaluation frameworks, embeddings & vector search, OpenAI, Anthropic, AWS Bedrock, Ollama, LangChain, Hugging Face

Languages — Python, TypeScript, Java, Rust, Go, C++, SQL

Frameworks — React, Next.js, FastAPI, Spring Boot, Node.js

Data & ML — PyTorch, scikit-learn, Spark, Kafka, Ray, MLflow, Triton

Data stores — PostgreSQL, MongoDB, Redis, DuckDB, SQLite, TimescaleDB

Infrastructure — AWS (S3, EC2, Bedrock, CloudWatch), Docker, Kubernetes, Terraform, GitHub Actions, Linux

Off the Clock

The Personal Column

A 35mm camera, city streets, and a standing weekend cricket match. Always curious, always building.

Open Source

Published Packages

benchtrace ↗

pip install benchtrace

Trace-first evaluation for LLM agents. One runtime-validated schema is written by the runner and read by grading, statistics, cost analysis, retrieval evaluation and reporting.

Pass-rate intervals use Wilson, with exact Clopper-Pearson available. A percentile bootstrap is deliberately not used, because for Bernoulli data it collapses at the boundary and reports 5/5 as a 95% interval of 100% to 100%. Two models are compared with an exact McNemar test, and it declines to name a winner when the interval includes zero.

PYPI · PYTHON · STATISTICS · 74 TESTS · MYPY STRICT

agentcassette ↗

npm install --save-dev agentcassette

Record an LLM agent run once, replay it in CI with zero API calls and no non-determinism. Zero runtime dependencies.

Recording happens at the semantic layer rather than over HTTP — messages, tool calls and results, normalised across providers — so a cassette survives an SDK upgrade or a provider swap, which is exactly where HTTP-level recorders break. When a replay genuinely cannot match, the error names the diverging turn and gives a path-level diff, so the fix is visible without opening the cassette.

NPM · TYPESCRIPT · 0 DEPENDENCIES · VITEST · JEST

Notable Works

Selected Projects

Understudy ↗

An LLM learns a task in a real banking UI once; the successful run becomes a typed capability that replays deterministically with zero model calls. The model never sees HTML, only an accessibility tree, so it cannot reference a selector the replay engine could not resolve. Every proposed action passes the same policy engine that guards replay, and irreversible actions stop at a human handoff where control transfers as a lease.

TYPESCRIPT · AGENTIC · PLAYWRIGHT · POLICY ENGINE

Mandate ↗

The control layer between an autonomous purchasing agent and company money. An agent sources and negotiates a purchase, then eleven deterministic checks decide whether it may receive a payment instrument at all — a failed policy check creates no card. Rules are rows a finance team edits, not code; every decision is replayable against the snapshot it was judged on.

NEXT.JS · AGENTIC COMMERCE · POSTGRES · POLICY

EDGAR Lens ↗

An agentic AI assistant over SEC 10-K filings — a multi-round loop that self-scores citation precision via Jaccard token-overlap and re-tries retrieval until a quality bar is met; reproducible evaluation suite across 19 filings achieving 100% citation grounding with median 1-round convergence.

LLM EVAL · AGENTIC · NEXT.JS · FASTAPI · OLLAMA

Chest X-Ray Detection ↗

End-to-end MLOps for YOLO chest-X-ray detection — Ray distributed training, MLflow experiment tracking with hyperparameter sweeps, ONNX conversion, Triton serving with canary rollouts and automated retraining on Kubernetes via ArgoCD.

PYTORCH · RAY · MLFLOW · TRITON · ONNX · K8S

Eventify ↗

An event ticketing platform — concurrency-safe seat reservation under optimistic locking, QR check-in, role-based JWT auth and Stripe payments with webhook idempotency.

JAVA · SPRING BOOT · POSTGRES · STRIPE

DealForge ↗

A multi-agent negotiation platform — a buyer agent bargaining in parallel against ten autonomous LangGraph seller agents with distinct strategies; Langfuse observability and per-agent evals.

FASTAPI · LANGGRAPH · LANGFUSE · NEXT.JS

Fraud Guard ↗

Real-time risk detection over streaming transactions — Kafka ingestion, Spark feature engineering, XGBoost scoring with Evidently drift detection and auto-rollback on threshold breach; Locust-tested to 250 req/sec at p99 ≤660ms.

KAFKA · SPARK · XGBOOST · EVIDENTLY · K8S

BidWave ↗

Reactive real-time auction service — Spring WebFlux + R2DBC + Lettuce, with a single-pass Lua script for atomic bid placement eliminating the read-then-write race; Testcontainers integration tests prove exactly one winner under concurrent contention.

JAVA · SPRING WEBFLUX · REDIS LUA · R2DBC

StreamBridge ↗

A high-performance distributed message queue — async Rust broker core (Tokio, DashMap), Go REST API via cgo/FFI — benchmarked at 8.9M msg/sec with p99 latency 167–208ns via Criterion.

RUST · GO · TOKIO · CRITERION

Raft-KV ↗

A linearizable key-value store implementing Raft — leader election, log replication, snapshots — with safety and liveness machine-checked in TLA+.

RUST · TLA+

GeoFlux ↗

A geospatial database with custom R-tree indexing, WAL crash recovery, and a cost-based optimizer.

PYTHON · FLASK

AirWatch Analytics ↗

Real-time air-quality monitoring on TimescaleDB + PostGIS — spatial queries over compressed time-series hypertables with continuous aggregates.

TIMESCALEDB · POSTGIS

Around Campus & Beyond

Service, Honors & Research

Leadership & Service

At NYU Tandon

  • Graduate Orientation Captain (Fall 2025) — planned the orientation ceremony and managed a team of 30+ leaders for the incoming graduate cohort.
  • Graduate Student Mentor (Summer 2025) — guided incoming CS & ECE graduate students through housing, NYC, and internships.
  • Facilitator, Building Belonging (Fall 2025) — led conversations on inclusion and community on campus.
  • Orientation Leader & Inclusive Dialogue Institute (2025) — campus tours, course guidance, and inclusive-leadership training.
Honors

Awards

  • International Student Leadership Award — NYU Office of Global Services, 2026. Nominated by faculty for contributions to improving campus life for international students.
Research

Publications

Solar Panel Damage Detection and Localization of Thermal Images

A YOLO-based thermal-imaging framework achieving 69% mAP for fault detection across 1,500+ curated images. read the paper →

Opinion

Why I Don't Trust a Demo

By Om Thakur

Most AI software looks great in a demo. Ten curated questions, ten good answers, a room full of nodding. The interesting question is the thousandth query, from someone who does not know the magic phrasing, on an ordinary Tuesday, when the answer is confidently wrong and nobody notices.

That gap is where I have spent most of my attention. At NYU Langone I built a retrieval system for clinicians, and then I built the thing that graded it: ten models, six providers, 1,950 runs scored on faithfulness, cost, latency, and whether the citations actually supported what the answer claimed. It found failures that looked fine on screen. That is the half of the job I find genuinely interesting.

The same instinct shows up elsewhere in my work. I reach for tools like Rust and TLA+ not because they are fashionable, but because they force you to be honest about failure before it happens. Healthcare sharpened it further. When the data is clinical and the pipeline moves fifty terabytes, "mostly works" is not a category.

I am looking for a team building AI products for real users that takes the unglamorous half seriously: the data model, the evaluation, the failure path. If that sounds like yours, the classifieds below are open.

Classified · Situations Wanted

WANTED: A team worth building with.

Software engineer, available now for full-time roles in applied AI, full stack, and backend engineering, anywhere in the U.S. Authorized to work through June 2029 on F-1 STEM OPT, so no sponsorship is needed to start. Replies within one day; coffee chats warmly encouraged.

"Good systems should fail gracefully, and engineers should, too."
© 2026 The Om Thakur Times · Available across the United States