Bring one measurable problem. We build the test, builders compete on the same inputs, and you pay only when a solution clears the agreed bar.
Three challenges are live now with $1,800 in prizes. One completed round has public results and code.
Live boards
How it works
01
Tell us the input, the useful output, and what solving it is worth.
02
We agree the data, scoring, qualifying bar, and practical limits before launch.
03
Builders compete on the same evaluation. Only work that clears the bar can win.
If nothing clears the bar, there is no forced winner. The bounty rolls over or comes back.
Challenges
Enter a live contest, inspect a finished result, or help shape the data for the next one.
Build a local, private Wispr Flow-style dictation tool for mixed-language speech — top qualifier wins.
10 builders entered
For restaurants and cloud kitchens: answer hidden questions from CCTV-style footage about caps, handoff, bottlenecks, visible steps, and what was not visible.
2 scored entries
Write one Python function that beats Arnav on a fresh live market window — no finance background, no money, no API key.
Live participant board
Trading agents · Round 1 · closed Jul 2, 2026
QQQ ended the same window at -3.9%. Arnav's code is public and now runs as the benchmark every Round 2 entrant must beat.
Use or adapt this work
Need a version for your own workflow? Tell us the use case. We will confirm the licence, deployment, and support scope before quoting anything.
Discuss adapting it →These briefs still need representative inputs and reliable answers. Share data or register interest; the bounty and scoring will be published before builders compete.
Turn sports footage or event feeds into accurate scores and game-state updates.
Input needed: Sample footage or an event feed, plus trusted final scores and game states.
Estimate the odds of real-world events, explain the signal, and trade against a fixed benchmark.
Input needed: Real questions, source odds, decision dates, and unambiguous outcomes.
Match purchase orders, invoices, and delivery records, then surface the exceptions that need action.
Input needed: Anonymized purchase orders, invoices, delivery records, and resolved exceptions.
Have a problem? Post it as a challenge and we will turn it into a scored test.
Live boards
Build a local, private Wispr Flow-style dictation tool for mixed-language speech — top qualifier wins.
10 entered; 8 have current scores; 2 have previous sample scores; 0 have qualified for the prize.
| # | Entry | Score out of 100 | Run status |
|---|---|---|---|
| — | Your engine here — fast, faithful Hindi+English · top qualifier wins $500 | ||
| ▸ | RambleFix · benchmark · current 8-clip check · Jul 24The current Hindi+English engine returned all eight finals; five clips were capped by final latency. | 63.66 | benchmark |
| 1 | Vishal · current 8-clip check · Jul 20All eight clips returned a final and this run cleared the current benchmark; two mixed-language clips remain the gap. | 82.41 | current check |
| 2 | Arnav · current 8-clip check · Jul 24The revised entry returned every final and is close to the benchmark, but three clips were capped. | 70.21 | current check |
| 3 | Harsimran · current 8-clip check · Jul 13All clips returned a final; mixed Hindi-English accuracy and slower finals remain the gap. | 59.54 | current check |
| 4 | Anmol · current 8-clip check · Jul 24The latest revision returned every final, but four clips were capped and it remains below the benchmark. | 59.25 | current check |
| 5 | Sham · current 8-clip check · Jul 13Two English clips beat the benchmark; mixed Hindi-English finals still missed too much. | 55.59 | current check |
| 6 | Rishchith · current 8-clip check · Jul 13Several finals were strong, but one clip failed and the slowest finals lost ground. | 52.89 | current check |
| 7 | Sankeerth · current 8-clip check · Jul 13Every clip finished, but several finals changed key meaning and arrived too slowly. | 45.76 | current check |
| 8 | Vishwas · current 8-clip check · Jul 20One clip returned a usable final; the other seven did not finish with usable text. | 6.25 | current check |
| — | Meet · previous 6-clip sample · old scoring · Jul 10Jul 13 returned no final; previous sample retained for context. | 41.90 | previous sample |
| — | Darshan · previous 6-clip sample · old scoring · Jul 10The output was blank or unrelated, and all six clips hit the time limit. | 20.00 | previous sample |
Watch · 2 minutes
A problem you keep meaning to deal with — and the move most people don't have a habit for yet: turn it into a challenge, let builders compete on the same test, and give qualifying builders a public proof of what they built.
After a published result, builders can claim their profile. Qualified builders who want a signed certificate can share that profile on LinkedIn with what they built and liked or learned; sharing never changes the score. Browse public results →
Deep dive: local dictation · ← all challenges
Implement draft(), push to GitHub, submit the URL. We run it offline on a fixed laptop. The leaders already beat the open-source baselines we tested; the prize bar is RambleFix on the hidden run.
# 1. fork → implement one function def transcribe(audio_path): # local model only; no network during scoring segments = stream_decode(audio_path) text = finalize_hinglish_mix(segments) return text.strip() # 2. push to GitHub → 3. email the repo URL $ git push && open mailto:submit@builderr.ai
Trading Round 2 · ← all challenges
Round 2 runs from the July 7 market open through the Sep 4, 2026 close. Two steps. Admission just checks your bot runs cleanly and respects the caps — a safety screen, not the ranking — then you're in. Then it trades scored forward-only — from the next open after submission on current market bars on real markets. Arnav, the Round 1 winner, runs from the same Round 2 start as the benchmark. Beat him to unlock the $1,000 prize pool and builder points. That live market window isthe score — in trading there's no truly unseen history(it's all public, and can be fit to), so the forward window is the only honest out-of-sample. Luck can't win it: each bot is scored only after it arrives, and the ranking is plain return over that window — and because admission caps leverage and how much goes into any one stock, it's not a who-gambles-most race. Full rules & why →
On Jun 15, the Nasdaq market tracker (QQQ) had recovered to +0.2%; Arnav was already +12.7%. By the July 2, 2026 close, it had rolled over to -3.9%; Arnav was still +5.9%. The winning code is public so builders can study what worked.
Admission also gives you a free read on how your bot behaved across three past market shocks. Here are two real bots we ran — same engine, very different results:
| SVB Mar 2023 | +12.19% | Sharpe 6.26 | MDD 5.5% |
| Q4 2022 rates | −3.52% | Sharpe −1.17 | MDD 11.9% |
| Aug 2024 carry | +2.81% | Sharpe 0.82 | MDD 13.0% |
Huge in recoveries, ugly in the rate downtrend. Admitted — not reckless, just bets on calm markets. Whether that wins depends on the forward window.
| SVB Mar 2023 | +2.13% | Sharpe 2.06 | MDD 3.9% |
| Q4 2022 rates | +3.38% | Sharpe 1.49 | MDD 5.5% |
| Aug 2024 carry | +1.61% | Sharpe 1.54 | MDD 4.5% |
Positive in all three regimes, every drawdown under 6%. All-weather. The bar to beat.
For fairness, every agent is scored forward-only — its $100,000 paper accountstarts at its first scored market session, and counts only from there. So no one can optimise against market history they'd already seen, and submitting later gives no edge (the “days live” on each row is its window so far). Same data and fills for everyone; account value, P&L, and trades are recomputed every market day by an open script (no hand-picking, no fakery). The winner is the best return over its live window (see the rules) — but Round 2 prize and points require beating Arnav. The board below is refreshed from the latest committed market run.
New here? Watch the 90-second intro — build a bot from a market thesis→Bring one AI or automation problem with an answer we can measure. We define the test, recruit builders, and run the competition. Here are a few examples:
If you can describe what goes in and what good looks like, we can usually turn it into a test in about a week, then run a few weeks of competition. The bounty is paid only when a solution clears the bar you approved.
How much should the bounty be? Size it to the problem. A simple way to think about it: if cracking it is worth a builder's weekend, the prize should feel worth that weekend. Around $1,000+ can work for a focused task; larger pools pull in stronger fields.
Send your problem →Two friends — product-and-tech geeks, both ex-founders — who hit this exact problem in our own work every week: you need an agent for something that matters, and no honest way to tell which approach is actually best. builderr is our fix.
We put real money on it. Each challenge has its own sponsor: Soham Sinha backs the trading bounty, Amit backs local dictation. Soham trades the winning bot on a live $100kof his own Nasdaq money after the contest, with the weekly P&L published publicly — a live ticker, from week one. No black box: you watch it work, or fail, in real time. That's the whole point — proof, not promises.
Current challenges have no platform fee. The bounty passes through to qualifying winners, minus agreed running costs. If no solution clears the bar, there is no forced winner. Have a problem we can score? Get in touch.