🤓 Latest Submissions
RadAgent: Grounded Radiology AI
RadAgent is a grounded multimodal radiology AI system designed to improve the reliability and transparency of AI-assisted clinical reporting. While modern Vision-Language Models can generate radiology reports directly from X-rays, they may also hallucinate findings or produce unsupported conclusions. In healthcare, where patient safety depends on accurate clinical decisions, explainability and trust are essential. RadAgent addresses this challenge through a multi-agent pipeline built for evidence-grounded reasoning. A specialist chest X-ray model first analyzes the image and predicts multiple thoracic pathologies using calibrated class-specific thresholds. Instead of passing raw predictions directly to a language model, validated findings are transformed into structured medical queries and enriched through a Retrieval-Augmented Generation (RAG) pipeline that retrieves trusted clinical knowledge. To improve interpretability, Grad-CAM visualizations highlight the image regions that influenced each prediction. A multimodal vision-language model then combines image features, calibrated findings, and retrieved evidence to generate a structured radiology report grounded in both image-derived information and external medical references. Reports can be exported as professional PDFs and translated into multiple languages while preserving clinical meaning. RadAgent is optimized for AMD Instinct™ MI300X using ROCm and vLLM, enabling efficient deployment of multiple AI components within a unified inference pipeline. More than a report generator, RadAgent demonstrates a practical approach to trustworthy medical AI—where every prediction is explainable, every report is evidence-backed, and every clinical conclusion is designed to earn clinicians' confidence.
13 Jul 2026
Verdict — AI counterparty due diligence agent
Verdict is an autonomous, event-driven counterparty due-diligence agent for business payments. Every B2B payment carries counterparty risk: shell companies, sanctioned entities, vendors with fraud histories. The evidence to catch them lives on the open web — but bot detection, CAPTCHAs and geo-blocks defeat ordinary automation, and raw AI output is not something you can gate money on. Different by design: Most web-research agents just summarize what they find. Verdict screens official OFAC sanctions lists BEFORE it trusts the web; it treats every scraped page as hostile input to prevent prompt injection; it verifies its own citations; and it remembers each counterparty so re-checks surface what CHANGED. It is not a research chatbot — it is an auditable control that shows its work. How it works: Give Verdict a company name. It (1) screens OFAC sanctions — a confirmed hit is an instant hard BLOCK; (2) reads the live web via Bright Data MCP (search_engine + scrape_as_markdown); (3) sanitizes scraped content; (4) returns a verdict — APPROVE / ESCALATE / BLOCK — with a 0-100 risk score and a cited source behind every finding. Single checks and batch screening that ranks a vendor list riskiest-first. Trust & explainability: Citation verification flags every finding as verified or unverified against collected evidence (anti-hallucination). A transparent risk calculation shows the score as a baseline plus signed contributions per factor. Auditor-style reasoning adds an evidence-confidence tag, a "why not higher risk?" block, and an adversarial/contradiction check. Each verdict exports an audit-grade PDF report. Tech: Bright Data MCP (live web, load-bearing), Groq Llama 3.3 with AI/ML API as an alternative provider, deployed on Hugging Face Spaces. Proven live: Apple → APPROVE. Wirecard AG → BLOCK 95/100, surfacing the real fraud, cited. Bank Melli Iran → instant OFAC sanctions hard-block.
31 May 2026
