A momentum stock-picker for the Nifty 500. 10 strategies running live, with a 46% 5-year backtest CAGR. The hard part was the data. NSE doesn’t publish historical index membership, so I scraped years of press releases to reconstruct it. Without that, a backtest only runs on the companies that are in the index today, which inflates returns.

github →

A weekly rebalanced allocator across Nifty large, mid and smallcap, a US index, gold, silver and debt. 25% CAGR over 14 years. Two versions run side by side. A pure quant model, and the same model with a Claude-powered macro overlay that can only reduce risk, never increase it. Running them in parallel shows whether the LLM is actually adding value or just adding cost.

github →

Built during my dad’s illness to read his scans more carefully. Runs on Google’s MedGemma 4B. Upload a CT, MRI or X-ray and it returns a structured radiology-style report plus a plain-English explanation, with prompts tuned per scan type. It can also line up two scans from different dates and flag changes in tumour size, position or appearance, which is normally a manual comparison.

github →

A Second Brain that reads everything I should be reading and remembers it for me, inspired by Karpathy’s post on LLM knowledge bases. I drop in a podcast transcript, Substack post, article or thread, or point it at an entire YouTube channel, and the agent reads it end to end. It creates or updates a page for every person, company, concept and topic mentioned, with each claim attributed to who said it, where and when. Over time each page gets dense with actual information, fully searchable and fully cited. Solves two real problems for me: getting a nuanced view on a topic without re-listening to five podcasts, and retaining what I consume instead of letting it evaporate within a week.

github →

Transcribes entire YouTube channels of 500 to 1000 hours at 150 to 200x realtime. A 150-minute video takes about a minute. Started as a way to build dense knowledge repos for niche podcast channels. Took me through speech-to-text at scale, from a single local GPU to parallel GPUs on Modal. Now feeds into the Knowledge Base above.

private

Always-on AI agents on a $5/month server. Jarvis routes tasks to specialists. Cadbury screens WhatsApp, Scoop sends a 9 AM retail digest, Postman summarises Substacks, Maharaj orders Swiggy, Mr. Market watches stocks, Archivist TLDRs YouTube and podcasts. Built on OpenClaw with Kimi K2.5.

private

Pulls AMFI data and quantifies the cost of picking Fund X over Fund Y. Simulates SIPs and lump sums, shows the gap in rupees, and does the same for direct vs regular plans so the distributor commission shows up as a real number over time.

github →

After India won the T20 World Cup my feed turned into nothing but cricket reels. Built a dedicated place to scroll through all of them. Hit 1000 users in 24 hours. Random reel feed, swipeable on mobile, with favourites and sharing.

github →

My first proper build that wasn’t on Lovable or Bolt, and the project that got me into Claude Code. Wordle for geography nerds. Daily puzzles on cities, countries and landmarks, with difficulty that scales by word length, a hint system, dark mode and shareable result cards. At peak 100 players used to play this, daily.

github →