Melbourne · Student · Builder
Engineering & business student at RMIT. Muay Thai, music, art, and code. Trying to build things that are honest enough to be trusted.
Hello
I'm a maker at heart — happiest when I'm learning something hard and turning it into something real that someone can actually use.
I'm studying BH111 at RMIT — a double degree in Electronic & Computer Systems Engineering (Honours) and Business. I chose both on purpose: I love understanding how a thing works all the way down to the hardware, and I care just as much about how it reaches people and earns their trust. The engineering teaches me to be precise; the business side keeps me honest about whether the work is actually useful to anyone.
Outside the screen, a few things keep me steady:
For what it's worth: I captained a cricket club to a championship, I'm a certified open-water diver, and I spent a couple of years in customer-facing work that taught me people come before product. I try to bring that same care into everything I build.
Why I build
Right now, AI is being trusted faster than it's being tested. That worries me — so it's where I've pointed my work.
Most of what I do comes down to one quiet idea: don't claim what you can't prove. My main project, falsify-eval, is a small open-source tool that catches AI systems which look right but aren't — and I gave it away under an open licence on purpose, so it can become a shared standard rather than something I sell. Good methodology helps more people when it's free.
The same care goes into the products. SiteIQ exists to take the spreadsheet-and-lost-message chaos out of running a small operations business — the kind of unglamorous tool that genuinely gives a small team their evenings back.
And the work isn't only for people who can pay for it. Through Bhardwaj & Sons — the family house I build under — I'm committing a standing share of everything I do to putting AI tools in the hands of non-profits and underserved communities, starting in India and Australia. The names and the audited figures get published. Nothing said that can't be verified — that's the whole point.
What I'm building
An open-source falsification harness that catches retrieval and ranking systems which look right but aren't. v0.1.6, on PyPI, Apache 2.0.
View on GitHub →A sovereign Sanskrit retrieval engine — 3-channel fusion (FAISS + BM25 + prosodic) grounded in Pāṇini's grammar, with >2× nDCG@5 over the baseline on Sanskrit benchmarks.
View on GitHub →An empirical study on structured-output adherence across model scales — and the prompting framework it produced. Small models follow schema precisely but hallucinate; larger models cite real papers but bend the format. Data published.
View on GitHub →A creator-intelligence tool that scores YouTube content via fuzzy logic, mines patterns behind the winners, then actively tries to falsify each pattern before trusting it. Surviving signals become ready-to-shoot content briefs.
View on GitHub →An autonomous Polymarket prediction trader using Hurst exponent regime detection and corrected half-Kelly bet sizing. Runs every 15 min, has daily-loss circuit-breakers, and ships a self-refreshing P&L dashboard. Default is dry-run.
View on GitHub →Run a commercial cleaning business from one screen — sites, contractors, clean records and risk alerts, all in sync. Live on Vercel with a real Supabase backend.
Open the app →I also collaborate with RMIT's Mathematics Department (under Prof. Lewi Stone), building learning resources and rebuilding the department's site so good teaching is easier to reach. Full proof of work →
Writing
I'm always up for a conversation about honest AI, low-resource languages, Muay Thai, or building something good. GitHub · sparshsharma219@gmail.com