Anson Boby

B.Tech CSE · CE Kallooppara · Kerala

I build thingsthat have to bedefensible

AI systems, on-device ML, native Linux tooling. Hostile data sources: a dataset with gaps, a device with no internet, a firmware that lies about its own fan speed.

See the workCopy my email
01 — selected work

Six things, no filler

6 projects
01
PYRO-0G screenshot

PYRO-0G

Will this material burn on the way to Mars?

NASA's microgravity combustion data turned into a live fire-safety cockpit: pick a crew atmosphere, pick a material, watch the flame regime change. Every record is dual-model verified — 80% pass, the rest are flagged in red rather than hidden. Built in 48 hours at NASA Space Apps 2026 with a team of six; I led it.

NASA PSIPython pipelineCanvasVanilla JS
02
~/twila

421M → 8.8B

ask
421M picks the effort
then
8.8B writes the answer

no api · local only

twila

A 421M model decides how hard the 8.8B model should think.

System 1 routes, System 2 reasons. "What's 2+2" gets 2k tokens and no thinking; an AIME problem gets 16k and full reasoning. Runs on a 15 GB laptop with no GPU, logs a receipt for every turn, and ships an eval that compares routed against one-size-fits-all. The 8.8B reasoner is TwIL-8B, which its model card benchmarks at 88–93% of frontier on several reasoning suites.

Pythonllama.cppRouting
03
PlateWise screenshot

PlateWise

Institutional kitchens overprepare by 30–45%.

Food-waste intelligence for hostels and canteens. Deterministic ML demand forecasting with recency decay, dual-unit auditing that never mixes kilograms with portions, and a surplus board with food-safety countdowns. 43 tests, no paid API.

TypeScriptForecastingTailwind
04
omarchy-localsend screenshot

omarchy-localsend

LocalSend, but it lives in your bar.

A Rust daemon on the official LocalSend protocol core with the GUI closed. Incoming transfers are never auto-accepted — an accept/decline card appears instead. Checksum-pinned binary, TLS certificate pinning, transfers that survive restarts.

RustQMLLinux
05
~/sadya-scan

sense → act

see
camera reads the leaf
decide
rules over model
do
servos drop the item

hardware in the loop

sadya-scan

Physical AI that inspects a banana leaf and fixes it.

Sense → think → act, with servos on the end. Gemini vision reads the sadya, a local Gemma reasons about what's missing, and dispensers physically correct the leaf. Hardware, models, and a full pipeline in one weekend.

PythonGeminiHardware
06
Vamanan GPT screenshot

Vamanan GPT

Onam's Mahabali, as an interactive storyteller.

A character-driven chat built around Kerala's Onam tradition: five scenes, a challenge, and a conversation that keeps the myth's voice. Next.js 16, React 19, Tailwind v4.

Next.jsReact 19Tailwind v4
02 — the rest

Also built

03 — about

Author's note

I'd rather ship the version that shows its 80% than the one that claims to be right.

Anson Boby, in flat editorial illustration
03 — profileKerala, 2026

I'm Anson, in my first year of a B.Tech in Computer Science at CE Kallooppara in Kerala, and first-year coordinator of the TinkerHub chapter there. Most of what I build starts with a source that won't cooperate, and my job is to make the answer hold up anyway.

At NASA Space Apps 2026 I led six people to PYRO-0G, a fire-safety cockpit over microgravity combustion data, where 80% of 25 records are dual-model verified and the rest are flagged rather than hidden. twila runs a 421M router that decides how hard an 8.8B model reasons, on a 15 GB laptop with no GPU.

First year. 26 public repos.

04 — contact

Say something

Open to internships, freelance, and anything where the problem is harder than the framework.

ansonboby@gmail.com
Copy my email