Local AI Platform

Running local models, RAG and agents usually means duct-taping together several tools with incompatible assumptions about state, auth and storage.

A single self-hosted platform that gives model serving, RAG and agent orchestration behind one consistent interface, so experiments don't each need their own throwaway infrastructure.

A model-serving layer for local LLMs sits behind a gateway API, alongside a vector store for RAG and an orchestration layer that exposes tools/agents through a consistent protocol, all self-hosted, no data leaving the network.

A running platform used to test agent designs, RAG pipelines and prompting strategies against local models before deciding what's worth building 'for real.'

Coding agents built most of the plumbing, the gateway API, the orchestration layer's boilerplate, deployment scripts, so time went into architecture decisions instead of wiring.

Chose to keep everything self-hosted and offline-capable from day one, even though it's slower to iterate than calling a hosted API, the point was learning what changes when you don't have that crutch.

Having one consistent tool/agent interface made it trivial to swap models and compare behavior without rewriting integration code each time.

Local model quality/speed tradeoffs were rougher than expected early on, agent patterns that work fine against frontier hosted models needed real rework to behave with smaller local ones.

Most of the hard problems in agentic systems are about permissions, context and observability, not the model itself.