Amartya Gaur

AI systems engineer. Agent orchestration, evals, infrastructure. Bengaluru, remote.

Founder and engineer at hunr.ai since August 2026. Before that, five years in Newfold Digital’s AI Center of Excellence.

[email protected] ·github.com/amartya-dev ·linkedin.com/in/amartya-gaur

Roles

hunr.ai (Kodaksh Private Limited) · Founder & engineer

August 2026 – present · Bengaluru

Technical screening that lets candidates use AI and then verifies they understand what shipped. Sole author.

Newfold Digital · Senior Software Developer, AI Center of Excellence

April 2021 – August 2026 · remote

Owned the agent orchestration framework behind the production support agents for Network Solutions and BigRock, the eval harness that gated their releases, and the generation stack behind the AI-native website builder. Earlier: roughly 100,000 Microsoft 365 accounts migrated to NCE.

PayCrunch · Founding consultant & technical architect

2022 – 2025 · part-time

Backend from zero for a UPI credit lending product at a Y Combinator-backed fintech. Hired and led the first three engineers.

Magicweave · Consulting

2026 · part-time

Advising on a game backend built by a friend. Wallets, currencies, stores and leaderboards.

Ohuru Tech · Freelance and client work

2020 – 2026 · alongside everything else

Client projects with the reusable parts published rather than copied between them: three packages on PyPI and a set of project cookiecutters.

Agent systems

hunr.ai · Founder · sole author · Since 2026 · Live
Screening that hands candidates a real repo, lets them use any agent, then checks they can defend it. Candidate code is graded on gen2 microVMs, not gVisor, chosen because only a real kernel gives you the in-job namespace unshare that drops egress mid-run, which then needs a shim to raise loopback again or every in-process test fails on principle. Per-challenge images are built from a package list with no daemon and no build step, because an agent-authored environment should not be able to run one.
When a routing benchmark can measure nothing · Ten pre-registered experiments · 2026 · Unpublished
I set out to validate a routing design and ended up with a result about the benchmarks instead: one can measure state-conditioned routing only so far as its state is not recoverable from its own transcript. MultiWOZ, SGD and ABCD each fail that. None of ABCD’s 1,004 test dialogues runs more than one workflow. Seven of ten registrations came back negative, including the one that killed my original hypothesis. Not written up yet.
Repo-to-docs retrieval · manu docs · 2026 · Prototype
The retrieval layer for a pipeline that turns a repository into documentation: a tree-sitter and personalized-PageRank repo map after Aider’s, an AST-chunked code index and a prose index over pgvector, each pinned to its own embedding model. Wired end to end, with no retrieval benchmark in it yet. I wrote the decision record surveying how Cursor and Cody do this before the code, built the repo-map third of it, and then wrote down which two thirds I had not.
A data layer that refuses · Built for an assessment · 2026 · Private
An analytics agent over a layer that will not return a figure whose unit is undefined, because one platform counts a view at one second and another at thirty. The brief expected the model to emit SQL; I gave it a typed scripting surface instead, precisely so it could not route around the guarantee, and wrote down that the sandbox is a semantics boundary and not a security one. The demo above is the idea. The build is not mine to publish.
ax, an agent asset registry · Internal, days old · 2026 · Unproven
Agents produce scripts and queries nobody can vouch for a month later. Each asset declares what it came from; when a source moves, everything downstream is marked suspect by graph traversal, with no model call anywhere in it, and gets back to active only by passing verification. The design doc argues mostly for deleting things, and ends with the condition that would prove the whole premise wrong.
Agent orchestration framework · Newfold Digital · 2024 – 2026 · In production
A sticky state machine with a workflow registry, carrying the production support agents for Network Solutions and BigRock at roughly a thousand conversations an hour.
Evals with a deploy gate · Newfold Digital · 2024 – 2026 · In production
739 regression cases, ~2,700 assertions, ~250 scoring metrics, wired into Jenkins as a blocking gate. Deterministic assertions can fail a build; model-judged ones report and stay out of the way.
AI-native website builder · Newfold Digital · 2023 – 2025 · In production
Prompt to live hosted site. LangGraph generation harness, sandbox orchestration, atomic publishes across OCI and Cloudflare. 149 of 334 commits; time to first preview down 48%.

Infrastructure

Public MCP server for deploys · Newfold Digital · 2026 · In progress
Ship to a managed VPS from Claude Code or Codex. Infrastructure, DevOps, the tool interface, and Cloudflare Enterprise endpoint authorisation.
Typesense, all four layers · Public repos · 2025 · Open source
A registry mirror, a Helm chart running a three-replica Raft cluster with per-pod DNS, an admin dashboard, and the indexer that feeds it. Hardened for OpenShift because it had to be.
Microsoft 365 to NCE · Newfold Digital · 2021 – 2023 · Shipped
About 100,000 accounts migrated, and the async systems underneath rebuilt to survive low-memory hosting.
Loyalty and rewards engine · Regulated fintech · 2026 · Private
Led engineering on a wallet and loyalty platform: a declarative campaign DSL with static checks and a pure evaluator, arbitration between rules that both want to fire, budget and liability ceilings, maker-checker governance, and an append-only audit chain, on regionally isolated deployment cells, because data residency is a deployment topology and not a config flag. Client work, so that is as specific as it gets.
Ohuru Tech packages · Freelance · 2020 – 2026 · Published
Client work with the reusable parts published instead of copied between projects: a FastAPI auth package with JWT, refresh, RBAC and field-level encryption; a Django-admin-style resource SDK for FastAPI; and a CLI that drops shared editor rules into a repo. Plus project cookiecutters for Django REST, React and Express that other people actually starred.
Magicweave · For a friend’s product · 2026 · Consulting
A backend for game developers: wallets, currencies, stores, leaderboards. I consult on it. I do not own it. The part worth stealing is the shape: Django used only as an ORM and never to serve a request, with one factory producing two FastAPI apps that have genuinely different trust boundaries, so the public SDK surface cannot accidentally grow an internal route.
UPI credit lending backend · PayCrunch · 2022 – 2025 · Shipped
The backend from zero for a YC-backed fintech: lending and payment rail integrations with underwriting on top, plus the Flutter client. Hired and led the first three engineers.
Jeet Kune Do Federation of India · Pro bono · one evening · 2026 · Not yet deployed
A Next.js and Postgres CMS to replace a federation’s compromised legacy site. Rate limiting is one atomic Postgres upsert because Cloud Run runs many instances; roles are read from the database on every request so revoking one takes effect immediately; and the branch seed skips any row a human has stamped verified_at, so re-running it cannot undo the person who phoned round.

Merged into other people’s repositories

Textualize/rich · Upstream · 57k stars · 2020 · Merged
Twenty-nine lines that let a download column render binary units. The smallest thing on this page and by far the most installed.
product-motion · OpenMontage · 49k stars · 2026 · Open upstream
A pipeline that makes product demo films out of the product’s own source. A deterministic scanner lifts the real design system out of a React, Next, Vue or Tailwind repository with every token citing the file and line it came from; nine directors then script, stage, animate and cut a film whose screens are replicas of the actual product, not somebody’s impression of it. Truthfulness is a chain here, not a promise: tokens cite lines, screens cite files, generated assets carry their sources, and a stage refuses to close without them. 4,316 lines across 39 files, open upstream. There is a film it cut, in the post about it.
Azure Speech for OpenMontage · OpenMontage · 49k stars · 2026 · Merged
Speech-to-text and text-to-speech as optional cloud providers, about 900 lines each, auto-discovered by the registry with no change to anyone’s existing pipeline. Answering the review was the better work: I had been requiring a config file from the repository under analysis, which is an arbitrary code execution path, and the shared cache key ignored schema defaults, so an omitted optional field and its own default hashed differently and paid generations ran twice.
neta-resume · Upstream · platform layer · 2026 · Merged
Two merged pull requests adding the platform layer to a public record of every Indian legislator: a portable ingestion control plane, and a Pulumi foundation whose config refuses to deploy a cluster whose API server would be reachable from the internet. Authored and CI-validated, never operated. The running site is somebody else’s stack, and I wrote that caveat into the README myself.
Creative Commons cccatalog · Upstream · now Openverse · 2020 · Merged
Test-on-pull-request CI, and tag blacklisting in the image store.

Writing

A pipeline that cannot invent your UI · Generated 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.
A container per job, without a daemon · Building 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.
Taking the network away from a Cloud Run job · Running 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.
What a routing benchmark cannot measure · I set out to prove a routing architecture and killed my own hypothesis on day one. What survived is a condition the standard dialogue corpora fail, including one where not a single test dialogue contains the decision being benchmarked.
Proving a challenge discriminates before anyone takes it · Every hunr challenge ships with two solutions: one an expert would write, one a plausible engineer would. If the hidden tests can't tell them apart, it never goes live.
Why a sticky orchestrator beat a smarter router · We kept trying to make the router smarter. What actually fixed support agents was refusing to re-route mid-conversation.
739 cases that can fail a build · An eval suite nobody can block a deploy with is a dashboard. Here's what it took to make ours a gate.

Earlier

Software engineering internships at American Express and Bruce Clay. B.Tech in computer science, Vellore Institute of Technology, 2021. Four WordCamp talks.

← back to the moving parts