🤝 Top Collaborators
.png&w=640&q=75)
🤓 Latest Submissions
BridgeCare
Discharge is the moment a patient goes home with written instructions: which medicines to take, how much, how often, for how long, when to come back, and which symptoms mean danger. Those sheets are prepared under time pressure from dense clinical documents, and a single miscopied dose or an unnoticed contradiction can harm a patient. Patients who do not read the document's language are at even higher risk. BridgeCare is a review assistant for nurses and discharge coordinators. An open model running on AMD GPUs (AMD Token Factory, MiMo-V2.6-Flash) drafts the patient sheet from the discharge document. The backend never trusts that draft: every field must cite the exact text it came from, medication values must appear verbatim in a verified quote (an altered "50" for "500" is caught), and a contradiction between two parts of the document blocks approval until a reviewer records a resolution note. The model cannot approve anything; only an authenticated reviewer approves one exact, hashed revision, and any edit or reprocessing invalidates that approval. The printable patient sheet is built only from validated fields and can be shown in Spanish through a deterministic, reviewable glossary that never translates medication names, doses or units and visibly flags anything it cannot translate. On five synthetic discharge documents, three consecutive real runs on AMD matched 142 of 144 expected medication fields, blocked the contradictory case every time, and never repeated an injected instruction. BridgeCare does not diagnose, treat, or calculate doses; it makes the last step of a discharge faster and safer, and it knows when to stop.
18 Oct 2026
CodeArchaeologist
PROBLEM Every mature company has a critical system nobody dares to touch. It keeps the business running, but its author is gone, it has no tests, and its business rules are buried in the code. When the board asks "How much will modernization cost, and where should we start?", the answer is usually a weeks-long consulting engagement or an unsupported opinion. SOLUTION CodeArchaeologist answers that question in minutes, with evidence. It takes a legacy repository (Python 3, Flask and SQLite in this version) and produces: 1. A verifiable technical dossier. IBM Bob audits the code in read-only mode and delegates in parallel to subagents for SQL, routes, security and dependencies. Every finding cites a file, line range and snippet, and a Python validator rejects any citation that does not exist literally in the code. 2. An auditable migration recommendation. A deterministic engine walks the call graph, SQL and complexity, and scores every route with a visible formula: value x testability x business data / risk. It returns the first cut, the route not to touch first, and a three-wave roadmap with PERT effort. 3. A board-ready DOCX memo with the decision, risk matrix, recommendation and traceability from every figure back to the code. On the controlled FacturaYa sample, it also runs a first Strangler Fig cut: the legacy route and its modern replacement pass the same characterization tests (6 of 6). TARGET USERS CTOs and architects inheriting critical systems, consultancies that need a defensible assessment before quoting, and boards that must approve modernization budgets. HOW THEY USE IT Users upload a repository as a ZIP, or open the ready-made FacturaYa analysis. They watch Bob work live, then browse the risks (each finding opens at the cited code), the call graph and the migration recommendation. They can ask Bob questions about the code from a side panel and download the memo for the board. In the Modernization Studio, they pick target technologies and Bob builds a step-by-step plan, applied only to an isolated copy after explicit confirmation. WHAT MAKES IT UNIQUE - AI proposes, code decides. Bob discovers and explains; reproducible code calculates risk, migration order and effort. If Bob contradicts the engine or invents a figure, its answer is rejected. - Evidence before narrative. In the public showcase, 12 findings are validated and all 14 citations pass. - Execution honesty. Every result is labeled live or imported (a real recorded Bob session). - Secure by design. The analyzed code is data, never instructions, and uploaded code never runs. HOW IT ADDRESSES THE PROBLEM Instead of weeks of manual review, a team gets in minutes a map of the risks, a justified starting point and an effort range, all traceable to the code. CodeArchaeologist does not replace the modernization team: it gives them the map, the order and the evidence to start on Monday.
27 Sep 2026
