🤝 Top Collaborators
.png&w=640&q=75)
🤓 Latest Submissions
StockShield AI
StockShield AI is an enterprise retail intelligence platform built for the AMD Developer Hackathon to eliminate global inventory distortions and stockout losses. The core problem in modern retail AI is the "LLM Hallucination Trap" — attempting to use large language models to compute inventory reorder math and time-series forecasts. StockShield AI solves this fundamentally by strictly separating mathematical computations from cognitive reasoning: Mathematical ML Engine: 100% of demand forecasting (HistGradientBoostingRegressor achieving R² = 0.995) and anomaly detection (IsolationForest) are handled by scikit-learn and supply chain inventory algebra, ensuring zero math hallucination. AMD Hardware Acceleration & AI Agent: Google Gemma 2 27B accelerated on AMD Instinct MI300X via ROCm and Fireworks AI interprets mathematical results, drafts C-level executive crisis memos, dispatches emergency supplier purchase orders, and generates overstock flash-sale campaigns. Automated Quality Gate: Enforces deterministic margin safety checks (discounted price > unit cost) before publishing promotional offers. Real-time Telemetry: Tracks token economics and business ROI multipliers down to fractions of a cent.
18 Oct 2026
Certo — Trust Layer for AI Agents
AI agents can access company data, call tools, update records and complete multi-step tasks. This creates value, but also serious risk: most teams still have no clear way to prove that an agent is secure, reliable and ready for production. Certo is a trust, security and optimization platform for AI agents. It helps AI startups, agencies and engineering teams test agents before deployment, identify failures and understand what needs to be fixed. Teams can connect an agent through an API, test environment or execution logs. Certo evaluates prompt injection, data leakage, unsafe tool use, hallucinations, instruction following, planning and response consistency. It looks beyond the final answer and checks tool calls, execution order and use of context. The audit combines deterministic checks, expected-output comparisons and model-based judges. The result is an explainable Trust Score supported by evidence. Each finding includes severity, failed behavior, the relevant trace and a recommended fix. Certo also shows a Potential Score, helping teams understand how much the agent could improve after fixing the highest-priority issues. This creates a practical loop: run an audit, inspect the evidence, apply fixes and test again. For the AMD Developer Hackathon, Fireworks AI is integrated into Certo’s evaluation pipeline as a model-based judge. It analyzes structured outputs and traces against a fixed rubric and returns a verdict used in the audit result. The MVP includes security and reliability probes, support for major agent frameworks and mappings to recognized AI security and governance standards. The demo shows a customer-support agent audit with a Trust Score, findings, evidence, fixes and a shareable report. Certo initially targets AI agencies and AI-agent startups that need to prove reliability to clients before deployment. The long-term goal is to become an independent trust layer for production AI agents through continuous testing and monitoring.
13 Jul 2026
