.png&w=256&q=75)
🤝 Top Collaborators
.png&w=640&q=75)
.png&w=640&q=75)
.png&w=640&q=75)
🤓 Latest Submissions
Bug Vaccine
PROBLEM Every team fixes bugs, and many of them come back. Under incident pressure, fixes ship without regression tests; "add a test" becomes a postmortem follow-up nobody finishes. New code, written by people who never saw last spring's incident, repeats the same mistake somewhere else: a missing null guard, float money math, an off-by-one. Coverage can't see this risk ("80% of lines executed" says nothing about whether you'd catch bugs you already shipped), and classic mutation testing is too slow and noisy for most teams. Each regression means paying for the same bug twice. SOLUTION Bug Vaccine treats a repo's own history as a vaccine. It mines bug-fix commits and postmortems; IBM Bob extracts the reusable pattern behind each bug (an "antigen") and flags follow-ups that were never done. Bob's parallel subagents, one per pattern, hunt for every place each bug could come back, including code written after the fix. Bug Vaccine then re-injects each historical bug one at a time and runs the real test suite: caught means immune; passing means the bug could return unnoticed. Bob in Agent mode writes "antibody" tests that guard each pattern, never touching production code, and a fresh Bob session that never saw those tests invents new variants as a blind check. Measured on our sample project: tests caught 2/6 past bugs (33%) before and 6/6 (100%) after; the blind check caught 6/6 unseen variants; 23 tests were added. TARGET USERS AND HOW THEY USE IT Engineering teams and maintainers. They run one command on their repo (or a GitHub Action on every pull request). Reviewers get a PR comment when new lines repeat a bug the company already fixed, with the team's own fix attached. "patch" applies those past fixes automatically, reruns the tests (rolling back if they fail) and re-scans; a pre-commit guard blocks the next copy, and Semgrep rules carry the knowledge into every IDE and CI. An offline dashboard shows the whole run, and a Debug lab answers "have we had this bug before?" for any crash, report or diff, with cited company sources. WHAT MAKES IT UNIQUE Mutation testing built from your own incidents rather than random mutants: few, meaningful mutants that each represent a real past failure. The core principle is "Bob proposes, the test runner decides": every number comes from actually running the tests, never from an AI estimate, and the blind held-out check guards against tests that memorise instead of generalise. WHY IT WORKS EFFICIENTLY A full check of six mutants takes seconds. There are no pip installs, and it works on Windows, macOS and Linux. It found what reviewers missed: the #57 fix had no test, and a module written after all three fixes repeated every one of them. In our PR demo it flagged 5 risks before merge and patched 3 repeated bugs with the team's own fixes (3 → 0).
27 Sep 2026
QuoteCheck
Procurement teams agree prices on supplier calls without knowing the market rate. A supplier names a number out loud, it anchors the negotiation, and by the time anyone verifies it the deal is done. Checking manually takes 15-20 minutes per line item, so on a 30-line quote nobody does it - buyers spot-check two or three lines and hope. The result is systematic, invisible overspend on routine purchases. QuoteCheck removes the delay. It takes a supplier negotiation call, extracts every part number, manufacturer, quantity and price the supplier stated, and pulls live distributor pricing for those exact parts through Bright Data across Octopart, DigiKey, Mouser and Arrow. Each line gets a severity-graded verdict and a suggested sentence to say back. The data layer is what makes the numbers defensible. The same part number is made by many manufacturers at very different quality tiers, so raw listings for one part can span 5x in price. QuoteCheck rejects manufacturer mismatches outright, trims statistical outliers, then takes the median of what survives - and shows the rejection count on every card. Where fewer than three comparable listings survive, it labels the figure indicative rather than guessing. Beyond price comparison it reads the negotiation itself. Hesitation detection flags filler words, hedging and evasion, because a price someone constructs on the spot is softer than one read off a sheet. Concession tracking measures how fast the supplier is moving and whether the slope has flattened. The honesty check says when a premium may be justified, because cheaper listings with six-week lead times are not a real option. And basket strategy reasons across all line items to recommend which to concede in order to win the one that matters most in absolute dollars. Built entirely on native.builder: React front end, Supabase with Row-Level Security, and two Deno edge functions so no API key ever reaches the browser
10 Aug 2026
