prakashchaudhary.tech — Developer & AI trainerCoding since 2022

Prakash
PROFESSIONAL BEGINNER · STILL LEARNING ·
DEV
Chaudhary

I know enough to build useful things and I'm curious enough to keep learning. History taught me patience. Tech taught me speed. This site is the evidence file.

Open the case files Show me receipts
Scroll — it gets more fun

// Manifesto

I build things.I break things.I ask why.

Show me a working app and I'll want to take it apart. Show me a headline about a 200-year-old revolution and I'll want to know the humidity that Tuesday. It's a minor personality disorder. But over four years, it managed to output some highly functioning, shipped code: drone autopilots that handle GPS failure better than I handle small talk, chatbots that remember things I forgot yesterday, and tools that catch fake news right before your uncle hits "send."

Patience
from old empires
Speed
from startups
Curiosity
chronic, untreated
Plan
usually none, works anyway
(01)

What I work with

// Languages & tools
01

Frontend

Interfaces that feel fast and don't fall apart. 3D in the browser when it earns its place, not because it's cool. (It is cool.)

  • TypeScript
  • React 19
  • Next.js
  • TanStack Start
  • Tailwind CSS
  • Vite
  • Three.js
  • GSAP
02

AI & ML

Models that run where they're needed — on a Raspberry Pi strapped to a drone, or entirely inside your browser tab.

  • Python
  • Gemma
  • XGBoost
  • LLaMA-2
  • Mistral
  • LoRA
  • mBERT / MuRIL
  • WebLLM
  • WebGPU
  • llama.cpp
03

Backend & data

APIs, auth, databases, push notifications, and the boring plumbing that makes the fun parts work.

  • FastAPI
  • Supabase
  • Firebase
  • PostgreSQL
  • Row-Level Security
  • Web Push
  • IndexedDB
  • Web Audio API
(02)

Receipts

// Real numbers, real trophies
prakashchaudhary.tech
Kathmandu, NP · est. 2022
Receipt №0001 · 2026
  • Hackathons entered0
  • Hackathons won0
  • Simulated Himalayan flights0
  • Crash-risk accuracy0.0%
  • On-drone re-plan time<0s
  • Manual river dives0
  • Servers PikuBot needs0
  • Fake-news models compared0
  • Years writing code0+
TotalWorth it
Sleep lost: not itemised
Thank you for scrolling
No refunds on curiosity
Kairos Engine — first place, Build With Gemma: Margadarshan
Exhibit A · 1st, MargadarshanWon
Team Pikachu — JunctionX Kathmandu overall winner
Exhibit B · JunctionX overallWon
(03)

Case files

// Two hackathons, two wins
№01 — Case file★ 1st Place — Gemma for Route Intelligence

The Kairos Engine

A Cognitive Autopilot for BVLOS Mountain Drone Delivery

Built for Build With Gemma: Margadarshan, winning first place (and $600) in the Gemma for Route Intelligence track. It gives a delivery drone something it has never had: judgment.

Prize won
$600

First place in the Gemma for Route Intelligence track came with a $600 cheque — the first time a side project paid for itself.

97.40%
XGBoost test accuracy
<2s
On-board re-planning
Pi 5
Runs fully on-drone
$600
Prize money

// Problem

Today's delivery drones fly like trains on a track — when wind or battery fails mid-flight, the only fallback is Return-to-Launch. In the Himalayas that move is often fatal, because home may sit behind a 4,000 m ridge the drone cannot clear. The Kairos Engine is not a flight controller; it is a decision-making brain that sits on top of one.

1 / 3
Gemma 4 E2BXGBoostPythonC++llama-cpp-pythonGGUFWeighted-DijkstraRaspberry Pi 5PX4/ArduPilot
Read the full Kaggle writeup

// Evidence 01 of 04 — tap a thumbnail

№02 — Case file★ Track & Overall Winner — JunctionX Kathmandu

Fluvius Predict

A live digital twin for hydropower river forecasting

Built for the JunctionX Kathmandu Hackathon. We won the track and the overall competition — and with it, the ticket to represent Nepal at Junction in Helsinki, Finland this November.

Representing Nepal
Finland

Winning JunctionX Kathmandu earned us a ticket to Junction in Helsinki this November — Team Pikachu goes international.

24h·7d·30d
Forecast horizons
0
Manual river dives
1st
Track & overall
Helsinki
Next: Junction, Finland

// Problem

Hydropower plants currently rely on manual measurements and decades-old data, leading to costly grid penalties.

1 / 3
AI ForecastingRodalo Hardware3D Digital TwinRain RadarSnowpack Data

// Evidence 01 of 04 — tap a thumbnail

(04)

The archive

// Hover to peek · click to open
8 files · 3 still being writtenFull archive on GitHub
(05)

How the work goes

// 4 steps, no magic
  1. 01
    Step 01 / 04

    Squint

    I stare at the problem and ask "wait, can this even be done?" Usually the answer is yes, with caveats.

  2. 02
    Step 02 / 04

    Read

    Papers, docs, other people's code, and — occasionally — a history book that has nothing to do with it.

  3. 03
    Step 03 / 04

    Ship

    Build the ugliest version that works. Then break it on purpose. Then fix it. Repeat until it stops breaking.

  4. 04
    Step 04 / 04

    Explain

    If I can't write it up clearly, I didn't understand it. Every project ends with a writeup or a demo.

// Then back to step 01. Always.