Partner at the desk.
AI in the loop every day: research, prototyping, writing, automation. Cloud for speed, a self hosted stack for anything private.
The stance
From idea to tested build.
Mid to high fidelity fast, sign off fast, then a clean handover so development and the proper Figma build run in parallel.
Rapid iteration
Mid to high fidelity screens, fast. Something real for stakeholders to react to.
Stakeholder sign off
Tight loops on the iterations. Decisions land because there's something concrete to point at.
Handover to development
Dev maps the solution and architecture while the patterns get refined alongside.
Hand built in Figma
Built from the ground up, correctly, so AI and automation can connect to it and test it later.
Dev architecture overlaps the Figma build, so the right foundation goes in without holding the timeline.
On Figma AI composers
Most get you halfway. The refactor for the other half costs more than building it right. The build that lasts is still the hand built one.
What I use, every day.
Six areas. Each one earns its keep on client work.
Agentic coding
AI in the IDE for multi-file edits, terminal work and test loops.
Tools
- Claude Code
- Cursor
- Repo prompts
- Sub agents
The big models
Frontier models for prose, planning, code and image work.
Models
- Claude Sonnet/Opus
- GPT-4 class
- Gemini Pro
- Model routing
Image & vision
Generative tools for moodboards and pipelines. Vision models for UI critique.
Pipelines
- Midjourney
- ComfyUI
- Vision API
- Screenshot critique
Local LLMs
Self-hosted models for work that can't leave my machine.
Stack
- Ollama
- Llama 3.x
- Qwen
- LM Studio
Automation
Workflow automation for the boring. File watching, reports, email triage.
Flows
- n8n
- Local agents
- Webhook pipes
- Cron jobs
Research & writing
Long-form research into structured notes. Drafts, edits, voice matching.
Uses
- Doc summarisation
- Transcript coding
- Voice matching
- Spec writing
Shipping product with AI inside.
What I think about when the model is part of the surface.
LLM in the loop UX
The model is one collaborator among humans. Design the handoff first, then the model behaviour.
Confidence & control
Two screens for one output: one when the model is sure, one when the user has to choose.
Streaming & tool use
Both change the grammar. The wait state becomes "here's what it's looking at".
Evals & observability
Prompts are software and they regress. Observability goes into the UI from day one.
Seven years of model waves.
ML and NLP at Gravicus, graph models at Chainalysis, LLMs every day since 2023.
Compliance models that surfaced risk from messy enterprise data.
On chain investigation tooling. Address clustering, transaction graphs.
Tool use shipped in Claude 2 and GPT-4. The daily workflow shifted.
Self hosted Llama and Qwen. The whole thing runs on my desk.
AI assisted surfaces in ArchCo's operational layer.
Let's design the real thing.
A workflow where the model sits beside the user? I'm the right call.