.png&w=256&q=75)
🤝 Top Collaborators
.png&w=640&q=75)
.png&w=640&q=75)
🤓 Latest Submissions
.png&w=828&q=75)
CommunityCare Voice Assistant
Billions of people cannot get basic health guidance or education because they live far from a clinic, cannot read, or cannot afford a modern setup. CommunityCare is a voice-first assistant built for them. A person speaks a question and hears a clear spoken answer. The answer also appears on screen, so it works with sound off. It has two modules. Health Triage listens to symptoms and says whether it is an emergency, something to see a doctor about soon, or manageable at home, with steps to take and warning signs to watch. It never diagnoses or names a condition, and every health answer ends with a reminder to see a real healthcare provider, even during a system failure. The Education module explains math, science, and language in short spoken sentences for low-literacy learners and ends with a follow-up question. Pipeline: AssemblyAI real-time speech-to-text, a Groq classifier that sorts questions into health, education, or off-topic, ChromaDB retrieval from a local knowledge base, a Groq model (gpt-oss-20b) that writes the answer, and Edge TTS to speak it. Off-topic questions are refused on purpose, because an assistant that answers everything cannot be trusted with health. API keys never reach the browser. The frontend is one HTML file, the backend is FastAPI, and the install is 338 MB. No login, no account needed. NGOs, clinics, and learning centers can run it on one shared device.
30 Sep 2026
.png&w=828&q=75)
TrustGate
AI coding tools now generate code faster than humans can review it safely. A developer can ship hundreds of lines in minutes, but speed creates new risks: hardcoded secrets, hallucinated packages, missing auth checks, prompt injection attacks, and spec violations that slip through human review undetected. TrustGate is an automated merge gate built to catch these AI-specific failure modes before code reaches production. When a pull request is submitted, five specialized IBM watsonx-powered checkers run in parallel on the code and deliver a clear PASS, REVIEW, or BLOCK verdict with exact cited evidence file path, line number, and quoted snippet before the code is allowed to merge. The five checkers are: Security Reviewer (finds hardcoded secrets, SQL injection, missing auth), Spec Conformance (reads the PRD and compares it against code using document understanding), Injection Scanner (finds prompt injection attempts in comments and README), Dependency Verifier (catches hallucinated PyPI packages), and License Checker (detects GPL violations in MIT projects). Every finding must cite exact evidence or it is not reported. This eliminates hallucinated false positives that make AI security tools untrustworthy. The verdict engine scores findings (high=5, medium=2, low=1) and forces a BLOCK on any single high-severity finding regardless of total score. TrustGate integrates directly into GitHub via a GitHub Action that fires on every pull request and posts the verdict as a PR comment. The Explainable Trust Dashboard shows judges and reviewers exactly why a PR was blocked, with a human-approval-with-reason override flow that demonstrates responsible AI deployment. Built in 48 hours for the IBM Bob 2.0 Hackathon.
27 Sep 2026
TariffTwin - Proof-Backed Landed Cost Twin
Tariff rates on imported goods are changing constantly a new Section 301 regime rolled out across 60+ countries on July 24, 2026 alone and small importers have no fast way to know their real cost. Most end up guessing or paying a customs broker just to find out, often losing money on decisions they could have made with better information. TariffTwin solves this with a simple, honest workflow. A user describes their shipment product, source country, and value and instantly sees their estimated landed cost, broken down into base duty and current additional tariffs, each backed by a real, dated source. The core feature is the country comparison: switch the source country and watch the cost recalculate live, so an importer can immediately see whether sourcing from Vietnam instead of China actually saves them money. Where TariffTwin stands apart is its honesty. When current tariff data can't be confirmed for a given product or country, it says so directly Unconfirmed rather than fabricating a number. That single design choice is the whole point: this isn't a chatbot guessing at trade policy, it's a tool built to be trusted. TariffTwin was built entirely in native.builder, using AI/ML API for product classification and cost reasoning, with tariff data sourced from official US trade references (USITC HTS, USTR notices). It targets a real, common decision small manufacturers and importers face today, in an industry most hackathon projects don't touch.
10 Aug 2026
