Navin Ghimire नवीन घिमिरेSOFTWARE ENGINEER / SF BAY AREA

Curiosity,
put to work.

I build the application, the system underneath it,
and the experiment that proves it works.

Explore the work SOFTWARE · AI · PHYSICAL THINGS

01 / Selected work

FROM THE CODEBASE TO THE WORKBENCH

Under the hood.

Production systems. Measurable results.

03

TAILORED BRANDS / 2023–2026

Making a commerce backend
legible to agents.

I wrapped nearly the entire API surface of a production Node backend-for-frontend as MCP tools, so coding agents could exercise the service directly in testing. Then I drove adoption across the engineering team.

NodeRedisMCP

3senior engineers adopted
the tooling

04

TAILORED BRANDS / GROWTH

Shipping the feature.
Testing the difference.

A guided filter and a video pill, shipped as server-side A/B tests. Variant assignment lived in the application layer, with real cache interaction and no client-side flicker. The impact was projected; the experiments were shipped.

Next.jsAB TastyConstructor

2features delivered
end to end

05

INDEPENDENT / LOCAL AI

From a song
to a finished film.

A music-video pipeline on one 16 GB GPU: generated keyframes, animation, voice, and edit. Models run in sequence; cuts follow the measured beat grid. Quality gates catch drift before it reaches the final sequence.

FLUX.2LTX-VideoPython / ffmpeg

6,368frames assembled
to plan

02 / The labEXPERIMENTS THAT EARNED THEIR DISK SPACE

Follow the
interesting question.

A few turn into projects. Many don’t.
These are the experiments worth keeping.12 NOTES FROM THE WORKBENCH

01

foundry

Local video generation with GPU admission control and quality gates on every take.

296 takes · 264 benchmark rows
02

hey-claw

An always-on voice agent, from wake word to spoken reply.

129 commits · runs at logon
03

cc-command-center

Agent racing with a panel of verifiers to judge the results.

6 scored runs
04

bun-vs-node

The same instrumented service, built twice to test the claims.

8 claims · 2 contradicted
05

gaea over MCP

Terrain generation from real Earth coordinates, exposed as agent tools.

188 nodes · 1,223 parameters
06

sahitya karaoke

A Nepali karaoke studio and a singing-voice detector fine-tune.

DistilHuBERT · 0.915 F1
07

comedy, measured

Laughter detection, shot grammar, and reaction timing in Nepali comedy.

262 shots · 22 automatic cuts
08

nepal-law

An append-only, hash-chained archive of Nepal’s federal law.

17,011 documents
09

3dprint-nepal

Satellite elevation turned into a four-foot, tiled relief map.

24 watertight STL tiles
10

the moat atlas

A study of the YC portfolio, as a book, web app, and ambient display.

6,186 companies · 799 pages
11

claude, on glasses

An agent on Ray-Ban Meta glasses, with credentials kept off the device.

Built APK · photographed live
12

dark-factory

A software factory that plans, builds, checks, and challenges its own work.

93 commits · self-authored PR merged
03 / The road hereनवीन घिमिरे

Hold the bar.
Build the thing.

I’m Navin, a software engineer in South San Francisco. My work moves between production commerce, AI tooling, procedural worlds, and things I can actually hold.

Before software: six years in the US Army as a CBRN specialist, with a last tour in Atsugi, Japan. I served part-time through a computer science degree at the University of Houston.

My working style is simple: make the system legible to the agent, then hold the bar yourself. The specification, the evidence, and the final judgment are still the work.

Current toolkit

TypeScript · Next.js · Node · Rust / wgpu
Python · MCP & agent tooling · CAD