Working demos and experiments
Hands-on builds showing AI concepts in practice — not slides, not theory. Each lab ships with source code, a live demo, and an honest writeup.
Good Tools, Bad Tools
The model never sees your code. It sees the declaration you wrote about it. Two agents get identical questions, identical functions, identical data, and the same model. The only thing that differs is how their tools are described. Ask a question and watch their execution diverge live, then read the recorded experiment that separates what a tool's name contributes from what its description contributes, with all 36 runs published including the five scenarios that refused to break.
Enterprise Hybrid RAG
See how an AI customer-service agent combines enterprise data, Vector RAG, ontology and Graph RAG to resolve complex distributor returns. A fictional battery company, structured transaction data, a small policy knowledge base, a lightweight knowledge graph, and deterministic refund rules, with a live evidence panel showing exactly what the agent retrieved and why.
SAP Sales Change Impact Assessment
Turn a synthetic SAP S/4HANA Explore workshop's As-Is and To-Be sales-order discussion into an evidence-linked Change Impact Assessment. A bounded Qwen agent uses allowlisted MCP context, deterministic validation, and an explicit business-review boundary.
Enterprise A2A: Purchase Order Status
See how a buyer's procurement agent asks a supplier's fulfillment agent for the latest purchase order status through the A2A protocol, without accessing the supplier's ERP system. Two independently deployed Google ADK agents, a live Agent Card, and a real-time execution trace over a genuine cross-service A2A call.
Why Are We Always Late?
A CFO-facing agent that diagnoses why a month-end close keeps running late, tracing the recurring cause through system evidence and attributed interview testimony, and predicts the next occurrence before it happens. Proposes a controlled corrective action with a real authorization boundary — approval never grants authority, and the tool re-checks the acting role after every approval, against a live Postgres audit trail.
What Is an Agent?
A beginner lab that explains what an AI agent actually is. Learn how models, tools, tool definitions, agent loops, memory, and the harness work through a simple enterprise leave assistant with exactly two tools. Ask it a question and watch every decision in a safe execution trace: which tool it picked, with what input, and what came back. Level: Beginner. Time: 10 to 15 minutes.
When Does Fine-Tuning Earn Its Keep?
The same conversation run live through three models, base BART, a fine-tuned BART, and Claude Haiku, side by side with cost and latency. A working demo plus a CIO decision framework: prompt first, RAG for missing facts, and fine-tune only for proprietary judgment, volume economics, or data that can't leave.
Procurement Exception Triage Agent
An enterprise agentic AI pilot on Claude Haiku. The agent investigates delayed POs, blocked invoices, and vendor risk through governed MCP tools — with a deterministic $25K approval hook, structured errors, human-review escalation, live agent-loop streaming, and per-run cost tracking (~half a cent per triage).
Structured Data Extraction Pipeline
A self-correcting structured output pipeline for enterprise contracts. Implements Pydantic validation schemas, self-healing retry loops, Message Batches API for 100+ documents, and a human-in-the-loop exception dashboard.
Configuring Claude Code for a Team Workflow
Project CLAUDE.md hierarchies, path-scoped rules, an isolated forked skill, and multi-server MCP config with credential expansion — with an interactive rule-matcher demo and a knowledge-check quiz.
MCP HTTP Transport & Tool Use
A live demo of the Model Context Protocol over HTTP. Claude uses MCP tools to manage a real file workspace — read, write, list, delete — with every tool call visible in the trace panel. Includes a deep-dive on MCP concepts.
Resolve: Autonomous Support Resolution on MCP
An AI agent that reconciles two siloed systems of record in real time — billing and access control — and shows every step of its reasoning in a live trace panel. Built on a real MCP client–server architecture.
Fine-tuning a Language Model: A Working Demo
Side-by-side comparison of base GPT-2 vs a domain fine-tuned model. Shows exactly what fine-tuning changes — and when RAG is the better call.