Client work · for HackerRank
DiscoverwhichAIisbestatcoding.
- Client
- HackerRank
- Our role
- Front-end build, full-stack features
- Status
- Live · launched on Product Hunt
The problem
Model benchmarks run on synthetic tests and crowd labels from people who never shipped the code. Nobody asks the only question that matters: would you merge this?
The idea
Put two models in a live arena on real tasks, let working developers vote, and turn every vote into training signal — public, transparent, developer-sourced.
What we shipped
Interface design by the HackerRank team. Our job was the engineering — the arena itself, the data loop behind it, and the surfaces around the vote.
The battle view
Two models generate side by side against the same task, problem statement always in frame. You vote for the code you’d actually ship.
Leaderboards that mean something
Rankings by language, task type and difficulty — so “best model” becomes “best at Python,” “best at SQL,” not one blurry number.
The data loop
Every vote carries context — language, task, difficulty, comments. That metadata becomes DPO training signal, not just a tally.
Beyond the arena
Annotation jobs for contributors and a purchase path for teams that want the evaluation data itself.
The build
A game that’s secretly an evaluation pipeline — it has to feel instant while a lot happens behind the vote.
Front end
- Next.js arena — live, side-by-side generation
- Streaming output with the task always in view
- A vote that takes one gesture, not a form
Data
- Every vote captured with its full context
- DPO-ready metadata per submission
- NestJS services behind the arena
Surfaces
- Public leaderboards — language, task, difficulty
- Annotation jobs for contributors
- A purchase path for the evaluation data
In the wild
Have something worth shipping?
We take products from brief to production.