AI Systems & Backend Engineer

Agents that ship.
Backends that hold.

Anirudh Negi

I build the parts of an AI system that are expensive to get wrong: tool contracts, failure semantics, and the guardrails that let agents ship to production safely.

Now

Backend Engineer

GoFundMe

Building

Negi Ventures

AI · Tech · Finance

See the work

Selected work

Things I built and run

UtilixLive

One tool registry exposed six ways, so an agent and a human call the same 236 utilities

TypeScriptNext.jsMCPREST APINode SDKPython SDK

KaryfyLive

Six systems that carry a job search from discovery to signed offer, with an agent behind each one

FastAPIReactPostgreSQLOpenSearchKubernetesClerk

PaperTrade ArenaLiveBeta

Live market data and a news pipeline feeding a paper trading terminal you can compete in

Next.jsMarket dataStreamingPostgres

Schema DriftLiveLabs

Grades every OpenAPI change by who it actually breaks, because request and response changes are not symmetric

TypeScriptOpenAPIVitestNo backend

Agent OpsLiveLabs

Replays agent runs as traces so you can see where the tokens, the retries and the human approvals actually went

TypeScriptObservabilityVitestNo backend

Local SEO AuditLiveLabs

Checks any small-business site for the handful of things that decide whether customers can find it and contact it

Next.jsSEOVitestLive fetch

Mortgage AtlasLiveLabs

Compares a 30 and a 15 year loan on what most calculators leave out: the deposit as money spent, PMI ending at 80% LTV, and costs that grow

ReactAmortisationVitestNo backend

ThrumLiveLabs

Twenty-seven instruments synthesised in the browser and locked to one key, recorded into loops and arranged into a song you can download

Web AudioDSPVitestNo backend

From the desk

Thoughts & field notes

100 posts on running AI coding agents in production. Guardrails, evaluation, MCP, and the things that break at 3am.

6 Jul 2026 · 8 min

SBOM Diff Gates for AI Coding Runners Before Dependency Drift Ships

A practical guide to SBOM diff gates for AI coding runners using lockfile-linked SBOM snapshots, severity-aware policy checks, provenance verification, and reviewer evidence so dependency-changing agent patches stay fast without turning into supply-chain blind spots.

AI Coding · Supply Chain

5 Jul 2026 · 6 min

Network Egress Policies for AI Coding Runners Without Breaking Package Installs

A practical guide to enforcing network egress policies for AI coding runners with allowlisted registries, DNS-aware controls, staged bootstrap lanes, and audit logs so agents cannot quietly exfiltrate data while still completing real builds and tests.

AI Coding · Runner Security

3 Jul 2026 · 6 min

SLOs and Burn-Rate Alerts for AI Coding Gateways Before Slow Turns Become Failed Runs

A practical guide to defining SLOs and burn-rate alerts for AI coding gateways with first-token latency, tool success budgets, lane-aware error budgets, and fallback policy so operators catch drift before slow turns become broken automation.

AI Coding · SLOs

Read all 100 posts

Know about me

Backend engineer, and a little bit of

I build the parts of an AI system that are expensive to get wrong: tool contracts, failure semantics, and the guardrails that let agents ship to production safely.

Outside of work I architect and operate live products built almost entirely through agentic workflows. I own the design and the reliability; the agents write most of the code.

Most of what I build starts as something that annoyed me at work: a migration that should have been safer, a contract that broke quietly, a tool that should have existed.

◆ planning · 4 files in scope
├ contracts/openapi.yaml
├ src/charges/handler.ts
├ migrations/0042_idempotency.sql
└ tests/charges.spec.ts
✓ schema gate · expand-contract, no destructive DDL
✓ contract diff · 0 breaking, 1 additive
! write scope · migrations/ requires approval
  approved by anirudh · 12s
✓ tests · 34 passed, diff-scoped
✓ patch applied · PR #218 opened
agent wrote the code. the gates decided what shipped.

8Products live

100Posts written

236Utilities shipped

191MCP tools

What I’m exploring

Graph data modelling for recommendation systems
Event-driven architecture and delivery guarantees
Autonomous agents that ship to production safely
MCP servers and LLM tool-calling design
Developer experience and tooling

Get in touch

Let’s build something that holds

Open to backend and distributed systems work, and to anything involving agents that have to run in production.

GitHub LinkedIn Negi Ventures

Agents that ship.Backends that hold.