Websites, web apps and AI agents.
I'm Palash Lalwani. I build software for small businesses and for myself. I also study law at NLU Meghalaya.
Services
What I build for clients
Web apps and sites
Marketing sites and web apps in React and Next.js, hosted on Cloudflare. Built to load fast and to be found in search.
AI agents and automation
Voice agents, MCP servers and workflow automation. I built Trove, an MCP server for academic research.
GladeForm
My web design studio for clinics and wellness practices. It takes new clients.
Work
Things I've built
Trove MCPMCP server
A fail-closed MCP server for academic research over OpenAlex, Semantic Scholar, arXiv, PubMed, Unpaywall and CORE.
MurmurmacOS app
Private voice writing for Apple Silicon Macs. Whisper runs on the device, and checks reject any rewrite that adds a name, number or path.
LoggedMobile app
A workout tracker for Android and iOS. No accounts, no servers, no analytics. Data stays in SQLite on the phone.
NearCareMobile app
An offline health assistant for rural India. Gemma 4 runs on the phone, so it needs no internet.
Voice agent widgetAI agent
One script tag adds a talking agent to any website. It runs on LiveKit, Deepgram, Groq and Cartesia.
Review automationWeb app
A review and private-feedback tool for German businesses. A QR code in the shop starts it, and the server enforces the funnel.
StashmacOS app
Hidden corner controls for macOS, with an iPhone companion that mirrors and drives them.
GrizzBearMobile app
A paper-trading app for NSE and MCX futures and options, with its own backend.
fearmapDeveloper tool
Reads a repo's git history and writes a plain-English map of the files that are dangerous to change.
GitFlixExperiment
Stores a 1080p, 60 fps video as Git history and plays it back.
SHAdeExperiment
Turns every commit hash into a painting and your repo history into a gallery.
Writing
Notes on data and systems
The Tests Were Green and the Job Was Gone
A live derivatives platform stopped force-closing intraday positions for seven weeks and nothing anywhere failed. The unit tests passed, the config was populated, the docstring still described the feature. A case study in the bugs that only live behind a green suite.
The Headline Win Rate Is Not the Result: Decomposing P&L
A trading bot went live at 75.3% win rate, almost exactly matching its backtest. That should have been the end of the story. Decomposing the aggregate found two things hiding underneath it: a subsystem quietly losing money, and an edge that was decaying day over day.
The 99% Trap: A Field Guide to Label Leakage
A 4-class risk model hit 99.4–99.9% accuracy. That number was the bug, not the result. A mechanistic breakdown of label leakage, the sneaky constructive kind, plus the honest 70.5% model underneath it.
Contact