.png&w=256&q=75)
🤓 Latest Submissions
PLAYBOOK SOAR: AI Agent Security & Response
Every AI agent deployed today has a critical flaw: nobody watches what they actually do. PLAYBOOK SOAR is the answer — a security platform that delivers complete AI agent incident response, forensics, and compliance in a single deterministic pipeline. Here's the gap nobody closed: Guardrails like SupraWall block single requests but leave you blind to the full attack. SOC platforms aren't built for AI. Governance tools give dashboards but zero runtime protection. PLAYBOOK SOAR closes the loop from detection to remediation without human latency. The Judge Layer is fully deterministic — zero LLM in enforcement. It renders verdicts (ALLOW / DENY / QUARANTINE / ESCALATE) based on immutable rule-based decisions in <5ms. Immune to 4 known LLM-judge bypass patterns, validated against 55 test vectors. No hallucinations. No prompt injection escapes. Pure deterministic enforcement. The 4-stage pipeline runs in under 40ms: DETECT → CLASSIFY/JUDGE → ENFORCE → FORENSICS. End-to-end incident creation with chain-of-custody audit trails and SHA-256 verified evidence packages — court-ready out of the box. What's demonstrable today: Agent Swarm Simulator running on Gemini 3.1 Flash Lite via Vertex AI with Misbehavior Mode (forces 100% malicious actions), Veea Lobster Trap DPI Live Feed with real-time interception, 16 incident types (AGT-FIN-001 through AGT-GOV-016), NIST Policy Builder with 6 industry templates and conflict detection, compliance mapping (EU AI Act Art. 9/15/73, NIST AI RMF Agentic Profile, SOC 2 Type II), and 1-click board-ready compliance reports. Why this wins: Others guard single requests. We respond to the full incident lifecycle — automatically, deterministically, and compliantly. Live demo: https://playbooksoar.aiproofofconcept.in | [email protected] / demo123 | github.com/shamuddin/playbook
19 May 2026
ClinSight: Agentic Multimodal Clinical AI
Every year, 795,00 Americans are harmed by delayed diagnosis in emergency departments. Preliminary chest X-ray review takes 30-60 minutes. Rural hospitals wait 4-24 hours for teleradiology. ClinSight is an open-source multimodal clinical intelligence system built entirely on AMD hardware. It ingests chest X-ray images, lab values, vitals, and triage notes simultaneously — then reasons across all modalities through a compiled LangGraph agent pipeline on AMD Instinct MI300X via ROCm 7.0 and vLLM. Architecture: 5 parent agents orchestrate 7 subagents (12 reasoning nodes). Coordinator validates input and runs pediatric safety gates. Radiologist analyzes X-rays via Qwen2.5-VL-7B. Lab Analyst detects critical values and correlates patterns. Safety runs 3 parallel checks (contradiction, hallucination guard, bias audit) with a merge node. Documenter produces deterministic ESI scoring, differential diagnosis, and structured reports. Dual-Model Stack: Qwen2.5-VL-7B-Instruct (vision, ~14GB) + Qwen3.5-35B-A3B MoE (reasoning, ~70GB) = ~99GB / 192GB HBM3. Both models served simultaneously via vLLM on ROCm 7.0 — impossible on H100 80GB without quantization. Live Evidence: 50-case pure CXR benchmark on real MI300X. Mean latency: 23.02s. All 50 cases live, zero cache. GPU utilization: 10-49%, power 231-263W. rocm-smi evidence captured at baseline, during, and post-inference. Safety & Rigor: Physician-in-the-loop by design. Pediatric gate blocks adult-trained recommendations for under-18 patients. Bias auditor stratifies by age and sex. ESI scoring is rules-based, never LLM-generated. Apache 2.0 license.
10 May 2026
