Skip to content

Sanmitra Nagaraj

I build the layer nobody sees until it breaks.

  • Data infrastructure
  • Distributed systems
  • Agentic tooling
  • 5+ years
  • Billions of ad events
  • 1 pager that stays quiet
000%
Loading the introduction sequence
001Manifesto

The best engineering is invisible. When a data platform handles billions of events a day without anyone noticing, that's the work.

Ad-tech data platforms move billions of events a day, and the interesting problems aren’t the features, They’re the Database node that OOMs at 221 GiB mid-merge, the tunnel that’s restarted 328 times because a floating IP was never bound to the interface, the secret that lives in exactly one region on the day that region degrades.

I’ve learned to distrust the happy path. Most of my best work is a timeout set correctly, a fallback that doesn’t silently mask a misconfiguration, or a rollback plan written before the upgrade started.

Right now I’m pointing that instinct at AI systems, Building agents that read our codebase, diagnose production failures, and propose fixes. Same job, new tools: make the thing reliable enough that someone can sleep.

Also

Hackathon wins

Open-source contributor

Civil Defence Warden

002Missions

Mission files

  • 01

    Agentic on-call

    A three-phase agent pipeline that turns ETL failure alerts into fixes. Agent 1 reads partner docs and the codebase to produce a root-cause report; Agent 2 recommends a fix; Agent 3 validates both and raises the PR. Explicit edge-case handling and a staged rollout keep humans in the loop where it matters.

    LLM agentsPythonClickHouseCI/CD
  • 02

    TACEE

    Pronounced Tay-see (Trends, Analysis, Content, Engagement Engine). ML model to provide insights for LinkedIn posts. Generates predictive analysis and recommends the best time to post.

    ML - classifierPythonThree-js
  • 03

    FRIDAY

    Back in 2022, Early LLM - Speech and text analytics - using slack API and Alexa. A chatgpt wrapper created well before APIs were released - used a hacky way to bypass security and use the model. :P

    pythonslack-apiearly LLM
  • 04

    Identity resolution pipeline

    Joined ad-server delivery logs against third-party demographic appends to build a unified audience view — dictionary-backed lookups, argMax tie-breaking for latest-record-wins, and materialized views over partitioned MergeTree tables.

    ClickHouseSQLData modeling
  • 05

    Open-source Java dataframes

    Contributions to dataframe tooling in the Java ecosystem.

    JavaOSS
003Toolkit

Skills

004Contact

Get in touch.