01Work & Projects
Decentralized real-time communication
Infrastructure · Real-time
The Terminal
A trading platform that works across every market. Crypto, commodities, equities, memecoins. On-chain, self-custodied, 24/7. I built the indexer that makes that possible. It parses Solana transactions as they happen, tracks every token from its first liquidity event, and ranks everything without ever touching OHLCV data. The hot path never hits a database. Background workers recalculate as new data arrives. Thousands of events per second, single node, zero loss.
Mobile · WebRTC
Huddle01 Meet
When I joined, there was no mobile team. No codebase, no roadmap. We needed a video conferencing app that could compete with Zoom, built on WebRTC technology that wasn't designed for this scale. I led the team and built the architecture. Extended React Native where it couldn't reach with custom native modules. Built Metal shaders for 120 FPS video rendering. Pushed the platform to 600+ participants in a single room. Took crash rates from 5% down to 0.01% on low-end Android. I rewrote the rendering pipeline twice to get there.
Mobile · Real-time Audio
FarHouse / House
Farcaster had no native audio spaces. Anyone who wanted live conversations had to leave the ecosystem. I built FarHouse to fix that. It was the first audio spaces app for Farcaster, built from scratch for iOS and Android. The hardest part: iOS kills WebRTC connections the second an app goes to the background. I found a way to keep audio alive. 50K downloads in the first few months. The app later became House.
AI-native startup, rapid product experimentation
Mobile · AI
Nintee Community App
Nintee shipped 3 to 4 production-grade prototypes every 2 months. Speed was the whole strategy. I built the Community app for iOS and Android using React Native. Designed a reusable UI system that accelerated every experiment after it. Built AI agentic flows directly into the app. Text-to-action interactions where you didn't chat with an AI, you just did what you needed to do. Smooth, fast, production-ready across platforms.
Mobile · Gaming
WordRally
We needed real user data to validate hypotheses. So I built a word game. Shipped it to markets where user acquisition was cheap, watched the retention curves, iterated fast. D1 retention hit 25%. Average playtime climbed from 90 seconds to 8 minutes. The game was simple. The data loop was the real product.
AI · Infrastructure
RAG & AI Systems
Before AI agents were a term anyone used, I was building RAG systems. Personalized document ingestion, embedding, retrieval, generation. No frameworks, no libraries, no playbook. Just figuring it out as the field was being invented. The system could ingest a user's documents and answer questions against them with surprising accuracy.
AI · Experimental
Gen-UI Agent
Every AI interaction at the time was a wall of text. I built an experimental agent that could generate interactive UI from a prompt. Predefined UI blocks, rule-based rendering. Instead of describing a list, the AI rendered a component. Instead of typing, you tapped and swiped. The AI felt visual instead of verbal. Ahead of its time, but it worked.
Side projects I ship for fun
Android · AI Agents
Phone for AI Agent
An Android app that gives AI agents access to a real phone. Uses AccessibilityService to tap, swipe, navigate, and read the UI tree. The agent connects via WebSocket and controls the device through a Node.js relay server. No brittle automation frameworks. Just a phone the AI can use.
Mobile · On-device LLM
Cactus LLM POC
POC that runs LLMs on-device using the Cactus library. Tool calling, memory, RAG. All local, no server. Built with Expo, uses native modules for PDF parsing and on-device inference.
02Blog
I write about what I build. Occasionally about why.
03AI Workflow
The stack
OpenCode runs everything. Codex sits on the side and tries to break whatever I build. Like that friend who points out every flaw in your idea before you start. I've started keeping a CLAUDE.md per project. Not a long one. Just the rules I'd otherwise repeat: stack choices, file conventions, which folders not to touch. The agent reads it before every task. Saves me from saying the same thing ten times.
Context juggling
Multiple agents on one project used to break constantly. They'd overwrite each other's work, lose the thread, make conflicting changes. Git worktrees fixed it. Each agent gets its own branch with its own working directory. They share the same CLAUDE.md for ground rules, but never touch the same files at once. Prompts are short. Goal, constraints, output format. If more context is needed, I let the agent ask instead of guessing what it might need.
Shipping loops
Pre-built skills are the biggest time saver. QA run, code review, deploy check, internet research through the browser daemon. Each one is a workflow I used to describe from scratch every time. Now I just call it. A feature that took a full day now fits in one focused session. Code, test, fix, deploy, verify. The agent runs the loops. I decide what ships.
04Who I am
I'm Darshan. I've been building software since before I knew what a career was. Most of my time goes into experimenting with AI agents and LLMs, but I don't limit myself to one layer of the stack.
I started at Nintee building AI agents when most people hadn't heard of LangChain. Then joined Huddle01 early, built the entire mobile app by myself, and watched it grow to tens of thousands of users. These days I lead the mobile team there and quietly ship my own things on the side.
I'm drawn to problems that feel slightly too hard. Not for the difficulty itself, but because that's where the interesting work is. Where you have to actually think, not just execute.
Outside of work I journal, I have a rooftop cafe I disappear to when I need to think, and I read more than I probably talk about.
05How I work
Ship first, refine after
I've built entire products by myself under real time pressure. That teaches you to make calls quickly, cut what doesn't matter, and keep moving. I'd rather have something real in front of users than something perfect in a doc.
Learn by building
Solana, ClickHouse, AI agents. None of these were things I knew deeply when I started. I read enough to get going, then build until I actually understand it. That's just how I work.
Small teams, real ownership
I've led small mobile teams and worked alone. I prefer environments where people take real ownership. Where you're not waiting on a ticket, you're deciding what to build next.
Full stack, genuinely
Mobile, web, backend, AI pipelines. I've worked across all of it. Not because I'm trying to be a generalist, but because the problems I find interesting don't stay in one layer.
06Books
The Almanack of Naval Ravikant
Wealth and happiness as systems, not goals. I keep coming back to this one.
Zero to One
The only startup book that changed how I actually think about competition.
The Hard Thing About Hard Things
No platitudes. Just the actual difficulty of building, written honestly.
The Cold Start Problem
Network effects are the only moat that matters. This book shows you how to build one.
The Mom Test
Short, dense, and it fixed how I talk to users. Should be required reading.
Deep Work
The rooftop cafe runs on this philosophy. Distraction is the real enemy.
07Contact
Let’s talk.
I don’t bite.
If you’re building something interesting and need someone who can own things end to end, I’d love to hear about it. Happy to talk shop about real-time systems, AI, or whatever you’re working through.