🤓 Latest Submissions
DejaBug
DejaBug: your team's past bugs, replayed as training cases. THE PROBLEM A new developer's hardest job is debugging code they did not write, and the first real practice often happens in production, under pressure. AI assistants add a new risk: when the AI writes the fix, the new hire skips the practice that builds the skill. Yet every repository already holds the missing course: hundreds of real bugs, each with a fix and a test that proves it, unused because turning one into an exercise by hand takes hours. In Cortex's 2024 survey, 72% of engineering leaders said new hires need over a month for their first 3 meaningful pull requests. OUR SOLUTION DejaBug turns a repository's bug-fix history into proven cases in a detective game. 1. Mine: find commits that fixed a bug and came with a test. 2. Prove: in parallel git worktrees, restore the code just before the fix and run the fix's own test. It must fail 3 out of 3 times, then pass on the real fix. Anything else is rejected with a reason. 3. Brief: IBM Granite on watsonx.ai, or IBM Bob Shell with our Deja Forger mode, writes a spoiler-free case file: symptoms, evidence, three hints, the lesson. 4. Play: the player debugs in IBM Bob with Deja Mentor, a read-only Bob mode that coaches with questions but cannot write the fix. 5. Debrief: XP, ranks, badges, and your time and diff next to the original team's. TARGET USERS Players are new hires and junior developers. Buyers are engineering managers at companies with 50 to 5,000 engineers. Open-source maintainers can publish free case libraries. HOW THEY USE IT Install DejaBug and paste any GitHub repository (Go or Python, with tests) into Settings. The Forge shows the proof live: red lights for failing runs, green for the pass. On the Case Board the player takes a case and gets a clean copy of the code with no later history. They fix it in IBM Bob with Deja Mentor and press Run the tests: the engine runs the certified test on their real files. Green means CASE CLOSED. WHAT MAKES IT UNIQUE Replaying real fixes is how AI models are benchmarked; DejaBug uses it to train people. Every case is proven by the project's own tests, and the mentor cannot write code, so the learning stays with the developer. RESULTS IBM/sarama: 2,889 commits scanned, 703 fixes found, 27 proven cases. IBM/python-sdk-core: 31 proven. google/uuid, never seen before, went from a pasted link to a proven case in about 90 seconds. 56 cases, none written by hand. BUSINESS VALUE The US expects about 106,100 developer and tester openings a year (BLS). DejaBug is free and open source; teams would pay for a shared case library, leaderboards and a manager's view. If it cuts ramp-up by two weeks (our assumption), a company hiring 20 engineers saves about $103K a year. Try it: https://ahammadshawki8.github.io/DejaBug/ Code: https://github.com/ahammadshawki8/DejaBug Video: https://youtu.be/AQhqBSi0cAY Pitch Deck: https://github.com/ahammadshawki8/DejaBug/blob/main/docs/submission/DejaBug%20Pitch%20Deck.pdf
27 Sep 2026
HalfSpread
Ask a trading agent what its edge is and it will tell you about the signal. Ask what it costs to get out and you get silence. We measured it. Thirteen tickers, same market, same minute. Crossing the bid-ask cost 0.52% of contract value on NVDA and 1.27% on AAPL. On contracts with 26 minutes to expiry: SPY 14.29%, QQQ 33.33%, IWM 100%. The only variable is time remaining. The spread is not a rounding error. Near expiry it is the trade. Paying it twice is optional: a contract that expires out of the money costs nothing to leave. HALFSPREAD is built backwards from that. It measures execution cost on every leg before an order exists, ranks verticals and condors by expected value after that cost, then holds to settlement instead of buying its way out. Not a hunch. Vilkov's 2026 0DTE study was corrected for charging the half-spread at a hundredth of its true size. Every net Sharpe flipped negative. The structures it invalidated are the ones most 0DTE agents still run. Underneath we solve Black-Scholes ourselves because Alpaca's feed returns zero greeks, recover index forwards from put-call parity because it publishes no spot for XSP or SPXW, and count breach probability from 668 measured sessions instead of assuming lognormal returns. A language model reads live headlines for event risk and may only shrink a position. That clamp is in code, not in a prompt. On the judged account: two spreads settled, $71.00 realised, $7.00 paid entering, $0.00 leaving. Closed conventionally the same trades return $61.00. Not paying to exit was 14% of the gain. Every figure above re-derives from an append-only journal with one command: no key, no account, no network. It runs in CI and fails the build when a number stops reproducing. When our dashboard once showed $316 that had not happened, the bug, the deletion and the reasoning went into the journal as a correction. Alpaca Trading API, CLI, MCP server, Market Data. Zero dependencies. Price a trade yourself against the live chain.
4 Sep 2026
