🤓 Latest Submissions
RIMPILOT – Your Bookkeeping Buddy
Informal vendors support their families day by day, but almost none keep books. They have no time, no accountant, and no app is made for someone serving customers with their hands full. So they know they sold 120 soles, but the till holds less, and they don't know why. It is rarely theft. It is lunch, the kids' bus fare, money they took for themselves and never wrote down. RIMPILOT fixes this with voice. You press one button and talk like you would to a friend. RIMPILOT records each sale, each business expense and each withdrawal for yourself, and before closing it asks the question nobody asks: did you take anything out of the till for yourself today? Then it reads your day back: sold 120, spent 25, took 25, you have 70 left. It is built on the AssemblyAI Voice Agent API: real-time speech in Spanish or English, turn detection, barge-in, and tool calls that write to the ledger. The math is done by our code, never by the model, so the numbers always add up. Every entry keeps the exact sentence that created it. Every ledger is private, and the API key never leaves the server. Nothing to download, nothing to learn. Try it in English or Spanish, without signing up, at rimpilot.onrender.com.
30 Sep 2026
GroundTruth Guard
THE PROBLEM AI coding assistants write code that sounds right. The most damaging failure is not bad syntax, it is plausible code that does not exist: a method on the wrong object, or a package that was never published. stripe.charges.createRefund() looks perfect, but it is not in the Stripe SDK. Worse, createRefund does exist in Stripe on a different resource (application fees), so a tool that only searches for the method name lets it through. Hallucinated package names are also a supply chain risk: attackers register the names models tend to invent ("slopsquatting"). Today these errors surface late, and most checks ask another model to review the first one, inheriting its weakness. OUR SOLUTION GroundTruth Guard checks AI-generated code against sources that cannot hallucinate, with four agents running in parallel: 1. Package registry: every import is checked live against npm or PyPI. 2. Real type definitions: using ts-morph, it resolves property to type to member in the SDK's installed .d.ts files. It knows charges is ChargesResource, which does not declare createRefund, and suggests refunds.create(). 3. Sandboxed execution: the code really runs in a Docker container with no network, a read-only filesystem and an unprivileged user. A TypeError proves a method does not exist; a blocked network call proves it does. 4. Security scan: hardcoded secrets, concatenated SQL and eval. Confirmed hallucinations are stored by signature in a team memory, so the same mistake is rejected in milliseconds next time, with any arguments. TARGET USERS - Developers who code with AI assistants like IBM Bob. - Tech leads who need to know which mistakes AI tools make in their codebase. - Platform and DevOps teams who want a deterministic gate before code ships. HOW THEY INTERACT WITH IT - Inside IBM Bob: a PreToolUse hook checks every file Bob is about to write. If it fails, the write is blocked, and a custom mode (GroundTruth Verifier) makes Bob read the report, apply the fix and write again. - Dashboard: paste code from any AI tool and see the four checks, the evidence and the fix. - Command line: verify-file.mjs checks a file and exits non-zero on failure, ready for CI. WHAT MAKES IT CREATIVE AND UNIQUE - It does not ask another AI. Every verdict comes from real evidence: the registry, the real types, real execution. - It is built with Bob and enforced on Bob. The custom mode cannot edit .bob/ or .groundtruth/, so Bob cannot disable its own guard. - The team memory turns every caught hallucination into shared knowledge. HOW IT ADDRESSES THE PROBLEM EFFECTIVELY AND EFFICIENTLY - Effective: the error is stopped before it reaches disk, not found in production. In our recorded demo, Bob was blocked on charges.createRefund, read the report and self-corrected to refunds.create in one step. - Efficient: known mistakes are rejected from memory in milliseconds; new code runs the four checks in parallel, so total time is the slowest check, not the sum.
27 Sep 2026
