Projects

Things I have built, most of them from the ground up: a database server, a coding agent, scaled backends, AI tools and one offline Android app.

01Systems2026

Radish

A Redis-compatible server written from scratch in Java.

Speaks RESP2 and RESP3, so redis-cli and normal Redis client libraries connect to it unchanged. A single Java NIO event loop serves every client, and a background thread handles AOF writes so the hot path never waits on disk.

~175K GET/s and ~144K SET/s on redis-benchmark (50 clients, AOF on)
Leader–replica replication, AOF and RDB persistence, pub/sub, transactions
No third-party dependencies, only the JDK

Java
Java NIO
RESP Protocol
TCP Sockets
AOF Persistence
Docker

Code

redis-benchmark → radish:6379

$ redis-benchmark -c 50 -n 100000 -q
PING   ~151,000 req/s   p50 0.151 ms
SET    ~144,000 req/s   p50 0.151 ms
GET    ~175,000 req/s   p50 0.151 ms
HSET   ~164,000 req/s   p50 0.151 ms
LPUSH  ~159,000 req/s   p50 0.151 ms
# AOF everysec, RDB snapshots and LRU eviction all on
$ redis-cli SET lang java
OK

02AI2026

Codemon

A coding agent that runs in your terminal.

Bring your own API key from about 185 providers. Codemon streams answers, shows every file edit as a diff before it touches disk, and asks before running anything risky. It ships as one binary built with Bun, so there is no Node or npm to install.

Permission gate sorts every tool call into read, write, bash or network
Sessions saved in SQLite: resume one, or roll back all of its file changes
MCP servers, sub-agents, plan mode, and a headless mode for CI

TypeScript
Bun
Ink
Vercel AI SDK
SQLite
MCP

Get the binary Code

codemon

$ codemon --plan
# read and grep only, every write is denied
$ codemon --sandbox docker
# bash runs inside a throwaway container
$ codemon --rewind
# restore files from the last session
$ codemon run "review the diff" --json
exit 0 finished · 2 turn budget hit · 3 denied

03Systems2026

TrimURL

A link shortener built to keep serving under heavy load.

A redirect reads Caffeine, then Valkey, then PostgreSQL, and pushes the click onto a Valkey stream, so no database write sits on the response path. A consumer group writes the clicks in batches. Runs on Kubernetes, with Helm charts for three environments.

Autoscales from 3 to 15 backend pods
Load-tested with k6 up to 4,000 concurrent users
Rate limits per user, password-locked links, click limits, hashed-IP analytics
Prometheus, Grafana and Loki for metrics and logs

Spring Boot
React
PostgreSQL
Valkey
Kubernetes
Helm
Grafana
K6

Code

grafana / JVM (Micrometer)

Grafana dashboard showing JVM memory and HTTP latency for the URL shortener backend

04Mobile2026

CropDoc

Point the phone at a leaf, get a diagnosis. No internet needed.

An Android app for farmers working where there is no data connection. An on-device MobileNetV2 model names the disease, then the app suggests a targeted treatment and works out the dose, so farmers spray less and spray the right thing.

38 conditions across 14 crops, fully offline
Offline AI assistant (SmolLM2 on llama.cpp) with voice in and out
11 languages, including Hindi, Bengali and Urdu
Built at NextStep Hacks 2026

React Native
Expo
TypeScript
TensorFlow Lite
llama.cpp
SQLite

Download APK Demo video Code

CropDoc home screen with camera, upload and batch scan buttons

CropDoc diagnosis: tomato mosaic virus at 99% confidence

CropDoc offline assistant answering a question

05AI2026

Agentic Support Desk

An AI support agent that knows when to hand off to a human.

Reads a customer's message from web chat, WhatsApp, email or a voice note, checks real order and payment data, and drafts a reply. A verify step checks the draft before it is sent. When confidence is low it escalates, with a summary and timeline ready for the human agent.

One LangGraph state machine behind four channels
Hybrid retrieval: pgvector plus full-text search in PostgreSQL
Checkpointed: a human can hand a ticket back and the AI picks up where it stopped
Runs on a laptop with local embeddings, no paid services needed

Python
FastAPI
LangGraph
PostgreSQL (pgvector)
Redis
React

Code

ticket lifecycle

channel   web · whatsapp · email · voice
classify  intent → tool group
retrieve  pgvector + full-text (hybrid)
act       typed tools: orders, payments, shipments
verify    grounded? policy-safe? complete?
decide    reply  |  escalate with handoff packet
# human replies, AI resumes from its checkpoint

More projects

foolscap

$ foolscap merge a.pdf b.pdf -o out.pdf
$ foolscap split report.pdf --pages 1-3,7 -o out/
$ foolscap compress scan.pdf --level screen -o small.pdf
$ foolscap ocr scan.pdf -o searchable.pdf
$ foolscap-gui report.pdf

Systems2026

Foolscap

A Linux PDF toolkit: one Rust core, a CLI and a GTK4 app. Merge, split, compress and OCR PDFs from the command line or a desktop app. All PDF logic lives in one core crate that never prints or exits, so the CLI and the GUI call exactly the same functions.

Rust
GTK4
MuPDF
Tesseract

Code

brutshop / products

Full Stack2025

BrutShop

A neo-brutalist store with real payments. Full-stack shop with role-based access for buyers and admins, Razorpay checkout and live wishlist notifications. Shipped as one container with CI/CD to Docker Hub.

Spring Boot
React
PostgreSQL
Redis
Cloudinary
Docker
Razorpay

Live demo Code

Free-tier host, first load can take ~30s

socket-chat-nine-tau.vercel.app

Full Stack2025

ChatApp

Real-time chat with voice and video calls. Private and group chat over Socket.IO, with WebRTC calls, media sharing, reactions and Gemini smart replies. Messages are cached in IndexedDB, so a return visit loads from disk first.

React
Express.js
MongoDB
Socket.IO
WebRTC
Zustand
IndexedDB

Live demo Code

syllab-ai-sarcastic-soul.vercel.app

AI2025

SyllabAI

Turns a topic or a PDF into a full course. Generates structured courses, quizzes, cheat sheets and flashcards with Gemini, using Next.js Server Actions and PostgreSQL. Tracks progress, streaks and accuracy per topic.

Next.js
PostgreSQL (Neon)
Drizzle ORM
Clerk
Gemini AI

Live demo Code

guide weave pipeline

query     "how do I reset the router?"
retrieve  manual text    → ChromaDB
retrieve  manual images  → ChromaDB
generate  local LLM
guide     step-by-step, with figures

AI2025

Guide Weave

Visual how-to guides generated from product manuals. Built during the Samsung PRISM R&D internship. A multimodal RAG pipeline pulls both text and images from manuals out of ChromaDB, and a locally hosted LLM writes the guide. Custom benchmarks tuned retrieval accuracy and latency.

Python
Multimodal RAG
ChromaDB
Local LLM

Code