CyberSearch is an agentic security analysis system that evaluates every code commit using real-time external intelligence, contextual LLM reasoning, and persistent repository memory. Unlike traditional scanners that rely on static CVE databases, CyberSearch integrates live threat intelligence via Tavily, processes full commit context using IBM WatsonX, and maintains historical state with Redis to enable continuous learning across analyses. The system runs as an event-driven pipeline triggered by GitHub webhooks, performing multi-step reasoning to determine whether newly introduced changes are actually exposed to emerging vulnerabilities. Each analysis produces a risk score, classification, and actionable decision (allow, review, or block), along with a detailed explanation. By combining real-time data, agentic reasoning, and stateful memory, CyberSearch reduces detection latency from days to seconds and enables proactive security decisions directly inside the development workflow. The goal is to move application security from reactive scanning to continuous, intelligent, and context-aware defense. https://drive.google.com/file/d/1w7SfqttmU4J3LKmqV15LJOjURxP8-XgL/view?usp=sharing
Built at Enterprise Agents Hackathon: Powered by IBM Watsonx Orchestrate AI Labs, Tavily, and Redis
CyberSearch AI
CyberSearch is an agentic security analysis system that evaluates every code commit using real-time external intelligence, contextual LLM reasoning, and persistent repository memory. Unlike traditional scanners that rely on static CVE databases, CyberSearch integrates live threat intelligence via Tavily, processes full commit context using IBM WatsonX, and maintains historical state with Redis to enable continuous learning across analyses. The system runs as an event-driven pipeline triggered by GitHub webhooks, performing multi-step reasoning to determine whether newly introduced changes are actually exposed to emerging vulnerabilities. Each analysis produces a risk score, classification, and actionable decision (allow, review, or block), along with a detailed explanation. By combining real-time data, agentic reasoning, and stateful memory, CyberSearch reduces detection latency from days to seconds and enables proactive security decisions directly inside the development workflow. The goal is to move application security from reactive scanning to continuous, intelligent, and context-aware defense. https://drive.google.com/file/d/1w7SfqttmU4J3LKmqV15LJOjURxP8-XgL/view?usp=sharing
Keep exploring what builders shipped.
solosquard
mystery-shopper-AI
Automate the detection of signup friction using AI agents that navigate like real users and report issues instantly.
ContractIQ
ContractIQ
We built ContractIQ, a 6-agent procurement workflow that turns contract documents into a decision-ready renewal package. IBM watsonx Orchestrate manages the full agent flow, including routing, parallel execution, confidence checks, and approval gates. Redis powers workflow state, inter-agent messaging, audit history, and cached memory across the pipeline. Tavily brings in live external research like vendor pricing, market signals, and competitive alternatives to ground the final recommendation.
Midas
Midas
OpenClaw Finance Agent — Enterprise AI that learns your business without seeing your secrets. Finance teams want AI that understands their vendor patterns, approval workflows, and GL structures — but compliance (GDPR, SOX, CCPA) blocks them from exposing sensitive data to train it. The result is that companies only use AI for suggestions. It's too risky to let it actually execute anything. We built a local-first AI system that solves this with comprehensive tokenization. Before any AI model sees the data, every sensitive field is replaced with a structural token. Microsoft Corp becomes VENDOR_001[TIER_LARGE]. $150,000 becomes AMT_LARGE[CFO]. Employee emails become APPROVER_LEVEL_CFO. PII is never extracted in the first place — it doesn't enter the pipeline at all. The model learns your company's patterns from tokens alone. The system uses two-tier AI routing. A locally fine-tuned model (LoRA on Mistral 7B, 4-bit quantized) handles 80% of routine tasks like invoice processing, spend analysis, and GL posting — under 500ms at $0.01 per task. Complex reasoning like forecasting and risk analysis routes to Claude API at $0.50 per task. Smart routing decides based on task complexity and model confidence. Every AI decision passes through MCP governance servers before execution. Is the approver authorized for this amount? Is the GL account valid? Is there budget available? Are there any policy violations? All gates must pass before the system acts. Every decision is logged to an immutable, hash-chained audit trail stored in tokens — not real data. SOX compliant by design. The architecture runs two isolated AI agents: Hermes handles coding and infrastructure (Docker, GitHub, codebase intelligence, Doppler secrets), while Midas handles financial analysis and compliance (governance, audit, GL validation, Tavily research, Redis). Each agent has its own Claude Code CLI wrapper and dedicated MCP servers. Namespace isolation ensures Hermes never touches financial governance and Midas never touches infrastructure. Training improves continuously through three channels: Tavily batch-researches public financial data, the /dream command generates synthetic training examples on demand, and Honcho from Plastic Labs automatically optimizes prompts based on real usage patterns. The system gets smarter with every interaction. Custom AI trained on your tokenized data. Runs on a laptop. Zero cost per local query. Full compliance audit trail. Policy-governed execution. Built in 48 hours for the IBM Hackathon, Procurement and Finance Track. https://www.loom.com/share/d05ac73cb5f94fb1b2ad99df0e1d32ea