Anzal Abidi
A curious builder. Lead Forward Deployed Engineer at GUS.
Open to new work
I build things, then find out what breaks. Lately that means taking AI from “this could work” to something a business can run on.
I · The journey
From clones to systems that hold
2021
Copying
I learned to build by rebuilding things I admired, like Netflix, Hulu, LinkedIn and Hashnode.
- Jamia Millia IslamiaB.Tech Computer Science
Copying an interface teaches you what it does, never why it was built that way.
2022 to 2023
Proving
College projects that real people used, and a national hackathon win with Medusa AI.
- BEABRANDSDE Intern
- Medusa AISmart India Hackathon winner
Most interviews in this stretch were rejections. Writing them up was the useful part.
2024 to 2025
Shipping
At Zomato I started writing code that had to survive traffic, build pipelines and other engineers.
- ZomatoSDE-1
You design for performance early, or you pay for it later.
2025 to 2026
Leading
At Kenverse I led a team of 14 building Kai, and the job became keeping our decisions traceable.
- KenverseSenior Software Engineer, Team Lead
A team moves at the speed of its worst-documented decision.
Aug 2026 to now
Building
Leading forward deployed engineering at GUS, and building tools in the open for the part of AI nobody demos.
- GUS (current)Lead Forward Deployed Engineer
Getting AI to answer is easy. Knowing when to trust the answer is the work.
Every new beginning comes from some other beginning’s end.Seneca
II · Work
Systems I built and shipped
- Fig 1Web securityLook closerOn npmpurify-edgeRuns DOMPurify, unmodified, on a DOM I wrote for it in 800 lines, so HTML can be sanitized on Workers and edge runtimes, where jsdom cannot go.0 differences in 800,000 fuzzed comparisons
- Fig 2Agent provenanceLook closerOpen sourcewitnessA proxy between agents and model APIs that keeps a tamper-evident record of every call, with signed delegation and a hash-chained journal.+0.25 ms per call, 34k calls a second
- Fig 3ArchitectureLook closerOn npmfaultlineA living map of a codebase’s systems and the imports between them, which stops a change, or a coding agent, from crossing a boundary it should not.0 of 6 agent runs crossed the line, 3 of 6 without
- Fig 4Computer useLook closerOpen sourcefasthandsLets any model use a computer: it shows the model only what changed on screen, and checks the page has not moved before every click.91% fewer tokens than screenshots
- Fig 5Voice agentsLook closerOn PyPItringWrite a voice agent once and run it on any of three voice runtimes, with the latency tricks that keep a call from going quiet built in.441 tests, three runtimes, one contract
- 5 systemsHow each one worksEach taken apart, with the problem it solves, the part that does the hard work, and the numbers that prove it.All projects
Also built
- guardiana judge inside an agent’s loop that stops dangerous actions
- sheetstreama million spreadsheet rows in 82 MB, in Rust for Node
- Quicky Wikidocumentation compiled into a knowledge graph that dates its claims
- voilanarrated product demos that record themselves
- samesmella fly’s brain reacting to smell, in the browser
- AgentReplayreplay any agent run against a change
- AI System Designthe unglamorous half of production AI
- UnClaudean open-source coding agent for any model
- Medusa AIexam proctoring, Smart India Hackathon 2023 winner
- Espressocontributor to Zomato’s open-source PDF library
Instruments
- TypeScript
- Go
- Python
- Rust
- React
- Next.js
- Three.js
- Node.js
- FastAPI
- gRPC
- WebRTC
- PostgreSQL
- Redis
- Qdrant
- Docker
- Kubernetes
- AWS
- MCP
- Claude
- Gemini
- OpenTelemetry
III · Writing
Notes from the build
Off the clock
On rotation
On repeat
Sirens
Ludwig Göransson · The Odyssey, 2026
IV · Contact