Skip to content

I build AI systems and teach the internet how they work.

Senior engineer shipping LLM systems to production. I write up what they cost and where they break.

What I've built, and what I've explained

Tap a note to jump to it.

Built

AI systems in production. The live ones you can try.

  • Masterji

    An AI startup coach. One task each morning, proof each evening.

    helpindiebuild.com
    How it worksMasterji's landing page
    Problem
    Builders plan more than they ship, and a chatbot coach will agree with anything.
    Built
    The loop at the top of this page, plus a cost ledger on every model call and an MCP server, so Claude can file proofs too.
    Outcome
    LLM cost is ₹2.70 per builder-day, the median over 174 production calls.

    Django, litellm, OpenTelemetry, MCP, Cloud Run

  • Skill Issue

    A daily game: which AI skill fires for this prompt?

    playskillissue.com
    How it works
    Problem
    AI agents pick a skill by its description, and it's hard to guess which one will fire.
    Built
    Rounds drawn from public SKILL.md files. An LLM writes the questions, a blind LLM judge reviews them, and only the ones it agrees with get published. The judge is calibrated against a hand-checked set.
    Outcome
    A daily job publishes new rounds by opening a pull request.

    Next.js, Upstash Redis, Postgres, GitHub Actions, OpenAI, Sarvam

  • Call pipeline

    Live call transcription and AI call outcomes for a US debt-collections client.

    How it works
    Cost
    ~5.5¢ → ~2¢ per call, after prompt caching
    Volume
    1,828 live calls in one day, for a client handling 10M a year

    Client work, anonymised. .NET, Azure Speech, Azure OpenAI, SQL Server, Python

  • AI support agent

    A RAG agent answering customer questions for Shopify brands, at DelightChat.

    How it works
    Problem
    Support teams at D2C brands answer the same questions all day on WhatsApp.
    Built
    I tech-led a two-person team. The agent retrieves from each brand's approved FAQs and product data in pgvector; vector matching replaced LLM-based FAQ matching.
    Outcome
    15k queries a month on the agent, 50k+ on the AI text tools, and $50k+ in new ARR.

    Django, Postgres + pgvector, Celery, OpenAI

Explained

Write-ups on how these systems work.

Other things I've built

Earlier, at IIT Jodhpur: Hybrid PageRank on the GPU (12× faster), Maze search visualiser.

About

Mahendra Singh Choudhary
  1. 2026 - nowSaxeCap Senior Software Engineer. LLM pipelines and backends for US clients.
  2. 2024 - 26DelightChat Senior Backend Engineer. Tech lead on the AI support agent.
  3. 2022 - 24Deloitte USI Software Engineer. Document classification across three clouds, a 100+ TB lossless migration.
  4. 2018 - 22IIT Jodhpur B.Tech, Computer Science. JEE Advanced AIR 2890.

Say hello

Open to talking about AI builds, collaborations and roles.