DSDavid Steedman
Menu

06Model orchestration · media · verification

The local AI platform

2025–2026Case study
Local AI Stack runtime-governance interface showing pinned versions, available updates and manual review states

The problem

Different AI models, interfaces and media tools are useful in different ways. The challenge was to make them work as one governed local environment rather than a collection of disconnected chats.

My role

Platform architect, interface designer, workflow engineer and verifier.

The platform around the models

The model is not the product.
The system around it is.

A useful local AI environment has to coordinate interfaces, engines, storage, safety and review. I designed those as one operating system for real project work.

01

Ways in

React dashboard · command line · MCP tools

Different interfaces, one control surface.

02

Control

FastAPI · durable jobs · GPU leases

Long-running work can stop, resume and leave a receipt.

03

Engines

llama.cpp · vLLM · ComfyUI · provider routes

The right engine is selected for the task, not for novelty.

04

Evidence

Manifests · hashes · licences · provenance · rollback

A result is not finished until its source and state are knowable.

1,559Python test functions and UI tests

134model and workflow manifests

19recorded architecture decisions

1governed environment instead of disconnected tools

Constraints

  1. 01Sensitive local material must stay controlled
  2. 02Models and providers change quickly
  3. 03Long-running work needs provenance and resumability
  4. 04Outputs must be checked before they enter a real project

What I built

  • A FastAPI control plane, command-line interface, React interface and MCP integration
  • Provider-aware routing across local text, vision, embedding, reranking and media models
  • Image, video and LoRA workflows with durable jobs and fixed-seed comparisons
  • Verified model install, update, rollback and uninstall with licences, hashes and immutable revisions
  • Security boundaries, encrypted local storage, provenance records and validation gates
  • Human review points for consequential decisions
Key decisionsHow the work moved
01

Local first

Private source material and durable project state remain under direct control wherever practical.

02

Use models as a panel

Different systems are assigned bounded roles, with disagreements treated as information rather than averaged away.

03

Require receipts

Important outputs retain sources, checks and handoff state so work can be audited and resumed.

Measurable proof

1,559

Python and UI test functions/tests in the current repositories

134

model and workflow manifests

19

recorded architecture decisions

Artefact gallerySanitised & provenance checked
Local AI Stack runtime table with pinned versions, latest versions and manual review status

The live runtime-governance view keeps pinned versions, upstream changes, staged updates and human-review states visible.

Local AI Stack control interface; privacy-safe live capture on 31 August 2026 · public
Decision Room interface resolving a five-issue local AI security architecture round

Decision Room used as an evidence-led architecture surface: five security decisions separated, challenged, locked and prepared for handoff.

David-controlled Decision Room application; privacy-safe live capture · sanitised

What the work taught

01AI becomes more useful when the surrounding system is opinionated.

02The fastest generation is not the fastest completed workflow.

03Verification and privacy are product features, not administrative chores.

Next case studyTechnology at scale