Amartya Gaur

Amartya Gaur

I build the layer AI agents run on.

Founder and engineer at hunr.ai

What has he actually run in production?How does he test agents?

The top tier is dashed because it is not mine. It is the part you are building.

Underneath the agent

Three layers, and the piece of work that stands as evidence for each.

AgentsWhatever you are building on top.
  1. OrchestrationRouting and state that survive a failed tool call.≈1,000 conversations an hour
  2. EvaluationRegression cases in the pipeline, so a prompt edit can fail a build.739 cases, blocking
  3. InfrastructureSandboxes, images and clusters for code nobody has read yet.microVMs, no egress

Six things I have built

Picked from twenty-four because they ran somewhere real, for long enough to teach me something. The rest of them — including the ones that did not work — are in the full index.

  1. hunr.aiLets candidates use AI, then checks they understood it.something to try
  2. Agent orchestration frameworkA thousand support conversations an hour, on a state machine.something to try
  3. Evals with a deploy gate739 cases that can fail a build.something to try
  4. AI-native website builderPrompt to hosted site. Time to preview down 48%.
  5. Loyalty and rewards engineA campaign DSL with an append-only audit chain.
  6. product-motionPoint it at a repo, get a product film that cannot lie about the UI.
All 24, with what state each is in

Working notes

How a few of these were built, what they cost, and the parts that were not in anybody’s documentation.

  1. August 2026Format-on-save, after the editorEditors format a file when a person saves it. Generators, shell commands and agents do not. I built onwrite around the harder question: how to repair that without putting the file at risk.
  2. 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.
  3. 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.
All of it

Say hello

Agents, evals, or the infrastructure underneath — I am always happy to talk about any of it.

[email protected]Let’s catch upCal.com · fifteen minutes