Amartya Gaur

AI systems engineerBengaluru · UTC+5:30

Amartya Gaur

I build the layer AI agents run on.

Five years of that was inside Newfold Digital’s AI Center of Excellence: a sticky orchestrator carrying about a thousand support conversations an hour, and 739 regression cases wired into the deploy pipeline so a prompt edit can fail a build. Since August 2026 it has been hunr.ai, which I write on my own.

I also spent about 230,000 model calls trying to disprove my own routing design. Seven of ten pre-registered experiments came back negative. That half is in this index too, because it is the more useful half.

The indexAsk this siteamartya@hunr.aiGitHubLinkedIn

Solid bars were the main thing at the time. Hollow ones ran alongside it.

Six years, five spans

Two companies, a consultancy and two products — and for most of those six years, more than one of them was running at once.

  1. hunr.aiFounder & engineerSince 2026
  2. Newfold DigitalSenior Software Developer, AI Center of Excellence5y 5mo
  3. PayCrunchFounding consultant, part-time2022 – 2025
  4. MagicweaveConsulting on a friend’s product2026
  5. Ohuru TechFreelance and client work2020 – 2026
Five entries have something you can operate on their page. Those are marked demo.

24 things, and what became of each

From a thousand support conversations an hour down to twenty-nine lines in somebody else’s repository. Every entry says what state it is actually in — in production, a prototype, or a notebook — because the reader who bothers to check is the one worth keeping.

  1. hunr.aiLets candidates use AI, then checks they understood it.Founder · sole authorSince 2026demoLive
  2. When a routing benchmark can measure nothingTen pre-registered experiments. Seven negative. Those were the result.Ten pre-registered experiments2026Unpublished
  3. Repo-to-docs retrievalA repo map, two indexes, and no benchmark yet.manu docs2026Prototype
  4. A data layer that refusesIt argues rather than hand you a wrong number.Built for an assessment2026demoPrivate
  5. ax, an agent asset registryMarks an asset suspect the moment its source moves.Internal, days old2026Unproven
  6. Agent orchestration frameworkA thousand support conversations an hour, on a state machine.Newfold Digital2024 – 2026demoIn production
  7. Evals with a deploy gate739 cases that can fail a build.Newfold Digital2024 – 2026demoIn production
  8. AI-native website builderPrompt to hosted site. Time to preview down 48%.Newfold Digital2023 – 2025In production
  9. Public MCP server for deploysShip to a managed VPS from Claude Code.Newfold Digital2026In progress
  10. Typesense, all four layersA mirror, a Helm chart, a dashboard, and the indexer.Public repos2025Open source
  11. Microsoft 365 to NCEAbout 100,000 accounts, moved.Newfold Digital2021 – 2023Shipped
  12. Loyalty and rewards engineA campaign DSL with an append-only audit chain.Regulated fintech2026Private
  13. Ohuru Tech packagesThree on PyPI, so the same thing is not retyped.Freelance2020 – 2026Published
  14. MagicweaveTwo API surfaces from one factory, different trust boundaries.For a friend’s product2026Consulting
  15. UPI credit lending backendA lending backend from zero, and the first three hires.PayCrunch2022 – 2025Shipped
  16. Jeet Kune Do Federation of IndiaSeeds that refuse to undo the person who phoned round.Pro bono · one evening2026Not yet deployed
  17. flutter-blocBLoC by worked example. 61 stars, 46 forks.Public repo · Hacktoberfest2020Open source
  18. Flutter and Django authentication, in three parts191 reactions across the series, and 35 comments answered.DEV Community2020Published
  19. Django, Vue and the rest of itA cookiecutter, a CRUD app, and turning any HTML template into a Django one.DEV Community2020 – 2022Published
  20. Textualize/richTwenty-nine lines. The most installed thing here.Upstream · 57k stars2020demoMerged
  21. product-motionPoint it at a repo, get a product film that cannot lie about the UI.OpenMontage · 49k stars2026Open upstream
  22. Azure Speech for OpenMontageTwo providers, merged by the maintainer.OpenMontage · 49k stars2026Merged
  23. neta-resumeThe platform layer for a public record of every legislator.Upstream · platform layer2026Merged
  24. Creative Commons cccatalogTest-on-pull-request CI, back when it was cccatalog.Upstream · now Openverse2020Merged

Earlier: American Express, Bruce Clay, a B.Tech from VIT, and four WordCamp talks I am contractually obliged to mention.

Each figure is drawn from that post’s own numbers, so no two are alike.

Working notes

Longer pieces on the things in the index: how they were built, what they cost, and the seven experiments that came back negative.

  1. August 2026A pipeline that cannot invent your UIGenerated product videos show software that does not exist. Fixing that is not a prompting problem. It means making every frame cite the line of code it came from, and refusing to finish when it cannot.
  2. August 2026A container per job, without a daemonBuilding a bespoke image for every unit of work, from a package list rather than a Dockerfile, and why having no build step at all is the security property, not a limitation.
  3. August 2026Taking the network away from a Cloud Run jobRunning untrusted code with no egress, on managed infrastructure, without a VM you have to operate. The trick is one prefix, and the reason it looks broken at first is that loopback starts down.
All of it

What’s breaking?

Agents, evals, or the platform underneath. Any of those three, and I will have seen the failure mode before.

amartya@hunr.ai