All projects
Every bet in one place. Each card is a pain I ran into and the answer I ended up shipping.
AI research moves faster than any person can read. Discovery isn't the bottleneck — filtering the flood into what actually matters to you is.
A personalized daily digest of arXiv papers, ranked around your interests and streamed to your inbox before the day starts.
Paper stamp cards get lost, smudged, and left in the wrong jacket. Small businesses lose the loyalty touchpoint entirely — and repeat customers lose the reward they earned.
Digital stamp cards that live in Apple Wallet and Google Wallet, right next to the boarding pass. The merchant scans, the customer taps, the reward always survives the laundry.
verdict-guard
GitHub Action that catches dishonest tests before they reach main
Visit github.com/Helmo21/verdict-guardAI assistants ship Playwright tests that pass without actually checking anything — a green toast, a waitForTimeout, no real verification. Coverage tools count assertions, not honesty.
A GitHub Action that scores every test for meaningful verification and blocks the merge when the score drops. Deterministic first, LLM only on the borderline — under three cents per pull request.
asia-fertility
The multilingual tokenizer tax, measured across 16 Asian languages
Visit doi.org/10.5281/zenodo.21069313github.com/Helmo21/asia-fertilityLLM tokenizers structurally overprice non-English languages by 3–12× on the same content. Most teams pay the tax without ever seeing it in a line item.
A reproducible measurement across 16 Asian languages that turns 'language difficulty' into a procurement decision — with every run pinned to config, prices, and tokenizer versions.