About

Words: 330

Reading Time: 2 mins

Hi, I’m Sajan Kumar. I build AI systems that go from prototype to production and keep working at scale. I’ve been the founding ML engineer at a YC startup, scaled DeFi platforms to 2M+ users, and now ship computer vision systems processing 10,000+ streams a day for state infrastructure.

How I got here

I fell into production ML at OurEye.ai (YC22) working on CCTV analytics for 500+ networks. When a thousand cameras started streaming, the only thing that mattered was whether the system stayed up. That’s where I learned: anyone can train a model—getting it to serve millions reliably is the real game. Since then I’ve bounced across surveillance, crypto, security automation, and transportation, always focused on making AI production-ready.

What I actually do

  • LLM systems: retrieval/rerankers, multi-agent workflows, eval loops, and guardrails.
  • Vision at scale: distributed inference, GPU utilization tuning, low-latency pipelines on mixed hardware.
  • Platform & ops: serving, CI/CD, IaC, observability, and reliability for fast shipping.
  • Product journeys: 0→1 and then 1→100—shipping, measuring, hardening.

Currently

  • Abstract Security — Senior ML Engineer (Oct 2024–Present)
    Leading LLM-powered automation and multi-agent workflows; processing 20TB+ security data and cutting triage from hours to minutes.
  • Purdue University (TASI) — Graduate Research Assistant (Jan 2025–Present)
    Optimizing statewide CCTV tracking and re-ID with transformer embeddings; 80%+ GPU utilization and ~40% lower latency on 10K+ streams/day.

Previously

  • WCG Clinical — Senior Data & Analytics Engineer (Aug 2024–Oct 2024) — Tuned GenAI/embeddings APIs and pipelines to cut latency and cloud costs.
  • CoinDCX — Senior Software Engineer (Oct 2022–Jul 2024) — Scaled DeFi wallet/microservices to 2M+ users; $50M+ daily volume; LLM-powered portfolio automation and fraud detection.
  • OurEye.ai (YC22) — Founding ML Engineer (Dec 2020–Sep 2022) — Real-time video analytics on edge hardware across 500+ CCTV networks.

Highlights

  • Automated 60–70% of security triage with LangGraph-based agent workflows on vLLM.
  • Statewide CCTV analytics: >80% GPU utilization, ~40% latency reduction on 10K+ streams/day.
  • Built microservices with 99.9% uptime handling $50M+ daily transactions.

Reach me

Next: Resume →