What we're building in the open.
Experiments, open-source tools, and technical write-ups on MCP, agents, and applied AI. This is how we think — and what we ship.
Experiments logged
Open-source repos
Technical write-ups
We built a custom MCP server that queries live restaurant data for an agent.
A walkthrough of the tool schema, auth boundary, and how the agent decides when to call it.
Write-ups & experiments.
Routing agent decisions across 3 models
How we fall back from Claude → GPT-4o → Gemini on tool failure.
Demand forecasting with sparse data
What we learned forecasting for outlets with <90 days of history.
Claude as a tool-router, not a chatbot
Treating Claude as the decision layer above MCP tools.
Structured outputs for agent pipelines
Why we moved every internal agent to JSON-schema outputs.
Gemini vs GPT-4o on long-context RAG
A practical benchmark on our own document sets.
Multi-tenant AI without cross-talk
Isolating context and memory per tenant in a shared agent layer.
Open-source tools.
Open source is how an emerging studio earns global trust — so we ship useful tools for free.
mcp-crm-connector
A reusable MCP connector for common CRM APIs.
agent-starter-kit
A minimal multi-model agent starter with fallback routing.
forecast-lite
A lightweight forecasting library for sparse operational data.