.png&w=256&q=75)
🤓 Latest Submissions
VeritasCore
VeritasCore is a real-time AI auditing engine that runs entirely on AMD Instinct MI300X via ROCm — no external inference API, no vendor lock-in, no per-token billing eating into the audit budget. Six autonomous LangGraph agents continuously red-team a deployed model: a Triage agent fingerprints the target, then parallel Bias, Adversarial, and Drift agents fire a library of 60+ adversarial probes — paired demographic-parity tests, prompt injection, encoding-based jailbreaks — all served locally through vLLM on MI300X. A Compliance Mapper cross-references every failure against regulatory clauses. A Meta-Agent tracks whether the probe library itself is losing effectiveness and synthesizes new adversarial cases from recent failures, so test coverage evolves instead of going stale like a static eval suite. The audit concludes with a cryptographically signed (Ed25519) certificate — tamper-proof, verifiable, ready for regulatory submission. What makes this an AMD stack story specifically: probe generation, target-model inference, and the embedding-based bias/drift scoring all run as concurrent vLLM workloads on a single MI300X instance. Running the full six-agent pipeline — including the self-evolving probe factory — at interactive latency on one GPU is only feasible because of MI300X's 192GB HBM3, which keeps the target model, the red-team model, and the embedding model resident simultaneously instead of swapping between them. Unlike static AI-safety scanners that run a fixed checklist once, VeritasCore treats auditing as a continuous, self-improving process — and does it without sending a single prompt to an external API.
13 Jul 2026
