Building AI for India's legal system.
An India-first AI workflow layer for litigation. Not chat-with-PDF — search, verify, then work.
Nine production modules: RAG-based case research, citation verification, AI drafting, document analysis, eCourts-integrated case management, evidence vault, advocate diary and litigation calendar.
Hybrid retrieval: BM25 keyword search combined with vector embeddings, semantic retrieval and RAG across 50,000+ Indian cases — with source provenance, paragraph-level citation validation and explicit hallucination-mitigation workflows.
Reliability is the product. Evaluation runs on citation accuracy, source provenance, retrieval relevance, stale-law detection and legal QA.
Validated the hard way: structured discovery with 40+ lawyers, associates and chambers; a 5-day diary study; a 30-lawyer concierge pilot; a 200+ waitlist.
Startup India / DPIIT recognised. ₹1.5 Cr pre-seed process led founder-first, on a 100+ investor CRM with automated personalised outreach.
Privacy-first local inference as a Python package, for environments where the data cannot leave the building.
Built and published for regulated and privacy-sensitive deployments — legal, health, public sector — where sending a prompt to someone else's cloud is not an option.
Distribution was the experiment: launched organically, reached ~50,000 Reddit views and 100+ upvotes, and crossed 100+ GitHub stars without a single paid impression.
Top 82 of ~43,000 applicants at Eureka!, IIT Bombay. Top 100 at Red Bull Basement.
Carried through two of the largest student-founder funnels in the country in the same year.
Eureka!, IIT Bombay — Top 82 from approximately 43,000 applicants. Red Bull Basement — Top 100.
A plug-and-play kit that makes a 20-year-old lathe legible: current, vibration, temperature, at the edge.
Low-cost retrofit combining current, vibration and temperature sensing with an edge gateway for anomaly detection, predictive maintenance and energy optimisation.
Shortlisted for MSME Idea Hackathon 6.0 and evaluated through SIIC, IIT Kanpur — with a proposed ₹15 lakh grant plus laboratory access for prototype and validation.
Four agents, four voices, four personas — running on local inference so the audio stays home.
A four-agent voice system with distinct personas and separate voice pipelines.
ElevenLabs speech infrastructure on the front, AirClaw local inference underneath, for privacy-oriented execution end to end.
A single-page cockpit orchestrating agents across business workflows. Packaged and sold.
React SPA that coordinates multiple AI agents across business workflows — and then, unfashionably, got packaged for digital distribution and shipped as a product.
A diagnostic and prep platform simulating GitHub's Agentic AI developer assessment.
Built and launched a full simulator for the GitHub Agentic AI developer assessment, with a structured two-tier commercial offering.
A student-led agency that reached ~$1,000 MRR while I was still attending lectures.
Founded and operated a creative agency serving B2B clients across real estate, fitness and F&B.
I owned the whole loop — sales, lead generation, pricing, delivery, client management, team operations — and reached approximately $1,000 MRR running acquisition and delivery alongside college.
Co-founded for Indian D2C beauty — creator rosters and outbound infrastructure.
Built creator rosters, outbound infrastructure and client-acquisition workflows for Indian D2C beauty brands.
~18,000 subscribers built on hooks, retention tests and no ad spend. Currently dark.
An independent content brand grown to approximately 18,000 subscribers through hook-driven short-form content, relentless retention experimentation and organic distribution.
It is also where I learned that the first three seconds decide everything — which turns out to be true of landing pages, cold emails and investor meetings.
The back catalogue is private for now. The channel and the audience are still there; a restart is a when, not an if.
A lawyer's problem is not that the document is unreadable. It is that the answer has to survive a courtroom.
So the interesting unit of work is not retrieval. It is retrieval plus verification plus the workflow the verified thing has to slot into — a cause list, a drafting deadline, a diary entry, an evidence bundle.
That is why JurixAI is built as search + verify + workflow rather than a chat box. Hybrid BM25 and vector retrieval over 50,000+ Indian cases finds the material. Paragraph-level citation validation and source provenance decide whether it is allowed to be used. The nine modules are where the work actually happens.
The barrier to legal-AI adoption in India was never information discovery. It was verification, contextualisation, citation checking, research rework, and integration with how chambers already run.
We ran a five-day diary study with eight practising lawyers. Not a survey — a diary, filled in at the end of each working day.
87.5% of them hit at least one research task that ran past sixty minutes. 62.5% were tracking cause lists by hand.
Those two numbers reordered the roadmap. An hour lost inside a research task is not a search problem, it is a verification problem — you find the case in four minutes and spend fifty-six deciding whether you can rely on it. And manual cause-list tracking is not a feature request, it is an admission that the calendar and the research live in different universes.
Everything shipped since points at those two facts.
Before writing the product, we offered hand-made research packs to thirty lawyers. Concierge, unglamorous, done by a human at night.
Ten said yes. 33% acceptance. Four of the ten came back and asked for another one — 40% repeat on accepted users.
That repeat number is the only reason to build software. Acceptance tells you the pitch works; repeat tells you the job is real. The 200+ waitlist came out of those conversations, not out of a launch.
If you cannot sell the manual version to thirty people, the automated version is a hobby.
AirClaw is a privacy-first local LLM inference package — a genuinely unsexy category. It got ~50,000 Reddit views, 100+ upvotes and 100+ GitHub stars with zero rupees of spend.
Two years of running an agency and a YouTube channel taught me the part most engineers skip: the launch post is part of the software. The first three seconds of a thumbnail, a README, a cold email and a pitch are the same three seconds.
Titan Media reached roughly $1,000 MRR because I did the selling. Raw & Blunt reached ~18,000 subscribers because I tested hooks instead of guessing them. Neither had a growth budget either.
nickzsche is a hand-rolled desktop in a browser tab. No UI kit, no window library, no animation library, no icon pack, no wallpaper images, no audio files.
Next.js App Router and React 19, TypeScript throughout, Tailwind v4 for tokens. The window manager — focus stack, drag, eight-way resize, edge snapping, minimise-to-tray — is about four hundred lines of pointer-event handling in one file.
Every app icon is SVG I drew by hand, which is why the set stays sharp at whatever size the dock magnifies it to. The wallpapers are layered CSS gradients and SVG. The terminal is a real interpreter with history and tab completion.
Music has no MP3s: all four stations are scheduled note by note into the Web Audio API at runtime, which is why the visualiser can read actual FFT data off the graph. Photos are the real ones, pulled from Instagram and committed to the repo so nothing depends on a signed CDN URL.
Everything you can read here comes from one file, lib/data.ts. Everything you can change — light or dark, accent, wallpaper, motion, sound — lives in System Settings and persists to localStorage.