Six projects, each one starting from a real problem instead of a technology. For each, I've written the problem, my approach, and what building it taught me — with source code and live demos throughout.
Plate I.
FitAI
AI-Powered Fitness Coach & Progress Tracker
Stack — JavaScript · React · Vite · Node.js · Express.js · Supabase · Google Gemini API
the plates follow in order of making
Spotted a common problem: people struggle with accountability and tracking their progress on their fitness journey.
Designed and built a full-stack AI fitness platform where users can log workouts, track progress, and get personalised coaching.
Added secure login, real-time data handling, and AI insights using the Google Gemini API.
Built a multi-provider AI cascade with health checks and automatic fallback — if Gemini goes down, it falls back to OpenRouter, and then to Groq — so the AI responses stay reliable even within free-tier limits.
Took the whole thing from the initial idea and problem discovery all the way to deployment and real use, and learned how products change based on user feedback.
Focused on features that actually help with consistency, motivation, and accountability instead of just tracking workout data.
Kept iterating on the platform based on testing and how people used it, to figure out what actually gives users value.
Got real experience balancing engineering decisions, user experience, and product priorities while shipping an AI app.
Saw an opportunity for a platform built around people actually swapping skills with each other instead of just passively sharing content.
Designed and built a prototype where users can find others by skill, show what they can teach, and connect to learn.
Used product thinking to validate the idea through fast prototyping and iterating on features.
Looked into ways to match users based on complementary skills, interests, and learning goals so they'd actually make useful connections.
Focused on making it easier to find mentors, collaborators, and peers in both local and online communities.
Designed the platform around actually participating and sharing knowledge, not just passively consuming content.
Used quick iteration and feature experiments to figure out which interactions actually gave users value.
Learned a lot about community-driven products, user engagement, and how hard it is to build network-based platforms.
Plate III.
LocalLens
Zero-Friction Local Discovery Platform
Stack — React · Vite · OpenStreetMap · Leaflet.js · Nominatim API · Overpass API
Noticed how annoying regular map apps are — you have to search manually, go through a bunch of screens, and deal with accounts before you can even find places near you.
Built a privacy-first local discovery app that instantly shows nearby restaurants, cafés, hotels, parks, attractions, and shops using real-time GPS.
Cut out a lot of the manual effort of normal map apps — no repeated searches, no switching categories, no typing in your location — you just open the app and immediately see what's around you.
Added automatic location updates and radius-based filtering so the recommendations stay relevant as you move around.
Built the whole thing on open map data using OpenStreetMap, Nominatim, and Overpass APIs — a fully serverless setup that costs nothing to run.
Kept it simple and fast by dropping accounts, tracking, social features, and extra UI clutter, to make a real "open and discover" experience.
Saw a few early startups tracking their competitors by hand — spreadsheets, bookmarked pricing pages, checking rival sites week after week — so I built MarketMind so they can look at their competitive landscape whenever they want.
Designed it to be a briefing, not a dashboard: you put in one URL, and in about 60 seconds it reads the product the way an analyst would.
Built automatic competitor discovery that maps out and ranks the whole competitive landscape from that one URL.
Set up continuous monitoring of thousands of public signals — read around the clock, so the market gets watched even when the team isn't looking.
Made sure raw events never hit the user directly: every signal comes already thought through, with why it matters and a suggested action attached.
Kept facts, inferences, and recommendations clearly separate — verified sources are labelled as verified, and AI inferences come with their own confidence scores.
Plate VI.
LODESTAR
Know What Your AI Actually Did — Independent Evidence for AI Coding Sessions
Kept running into the same doubt with AI coding agents like Claude Code and Cursor — after a long session the AI says "I changed these files and all the tests pass", and you basically just have to believe it.
Built LODESTAR, a small command-line tool that sits between you and the AI and quietly watches what actually happens on your machine — file creations, edits, deletions, every command it runs with its exit code, and the Git state.
Using it is simple: install it, initialise it once in your project, then launch your AI assistant through LODESTAR and work exactly like you normally would — when the session ends, one command gives you a full report of what really happened.
The report proves things the AI's own summary can't — which files truly changed, whether the tests genuinely passed, and even whether code got edited after the last successful test run, which is exactly the kind of thing you'd never catch by hand.
Inside, it's a one-way pipeline: every observed event becomes a standard record in a SQLite ledger, and each entry is hash-chained to the one before it — like a tiny blockchain — so if anyone edits the evidence later, the standalone verifier catches it instantly.
Deterministic rules then read the ledger and turn raw events into plain facts, and the same verified evidence is rendered three ways — a terminal report, a local dashboard, and a shareable HTML file — so every view always agrees, and you can even replay the whole session.
Made it fully local and offline — no accounts, no cloud, no telemetry — because a tool whose whole job is trust shouldn't ask you to trust a server too.
How to use it
Install Node.js 22.13 or newer, then check the version.
node --version
Install LODESTAR.
npm install -g @gshanmukha/lodestar
Check it installed properly.
lodestar --versionlodestar status
Inside your project folder, set up recording. You only do this once per project.
lodestar init
Launch your AI assistant through LODESTAR instead of on its own, then just work like you normally would.
lodestar claudelodestar run codexlodestar run aiderlodestar run <any-command>— use whichever matches your agent.
When the session's done, see what the AI actually did.
lodestar report
06 An open invitation
Have a problem worth building for?
I'm always up for talking about intelligent systems, research, startups, and building tech that actually matters.