AI automation and agents
Agents that do the repeated work, and checks that catch them.
I build AI automation and agents for work your team repeats: intake, research, drafting, sorting, publishing. Each step that a model performs is checked by code or by a second model, and a person signs off where a mistake would be expensive.
- Input documents
- Agent & tools
- Validation
- Human review
What I run myself
I run a multi-agent delivery system on Claude Code every day: 100+ role-specialised agents and workflows that take a product spec toward a tested Flutter and Supabase codebase.
- One run produced a 445-file app with 2,027 passing tests.
- The first full run taught the main lesson: the code was healthy, yet a new user could not finish onboarding. Verification was tied to tasks, not to the user's path. I rebuilt it around user-journey acceptance, and the next run found and fixed the broken path first.
- Generator and critic on different models: one writes, another scores, a third audits.
- ScalePilot, my landing-page product: a guided brief goes in, a pipeline builds, checks and publishes a designed page.
As of October 2026.
What is included
An automation project starts with the process as it runs today and ends with a pipeline your team can watch and stop.
Process map
Each step written down: input, decision, output, who checks it, what a mistake costs.
Agents and pipelines
Claude or GPT models where judgement is needed, plain code everywhere else. Versioned prompts with evaluations, so a prompt change is tested like a code change.
Checks
Schema validation at every boundary, a second model to score output, rules that park a result instead of publishing it.
Human review points
Approvals in the tools your team already uses, such as Slack or email, at the steps you choose.
Integrations
Your CRM, email, ad accounts, databases and internal APIs, through official APIs.
Cost and logs
Token spend and run history recorded, so you see what each run cost and what it did.
From manual process to pipeline
The pipeline is built one step at a time, and each step is measured before the next one is automated.
Map
We go through the process with real examples and pick the step with the most repeated work and the lowest risk.
First step automated
That step runs on real inputs with a person checking every output.
Measure
We compare output with what your team would have done and fix the misses.
Widen
Review moves to samples and exceptions only where results earn it. The next step follows.
Questions
What kind of work can an AI agent take over?
Work with clear inputs and a checkable result: sorting and routing requests, extracting data from documents, first drafts, research summaries, publishing to several channels. Decisions with legal or financial weight keep a person in the loop.
How do you stop an agent from making things up?
By not trusting a single model. Output is validated against a schema, scored by a second model, and checked against source data where it exists. Results that fail are parked for a person, not published.
Which models do you use?
Mostly Claude and GPT, picked per step. Using different models to generate and to check makes errors less likely to slip through.
Will it work with our existing tools?
Usually yes, through official APIs: email, Slack, CRMs, ad platforms, databases. I avoid scraping or automating screens where an API exists.
Where does our data go?
Only to the model providers and services the pipeline needs, and you choose them. Data stays in your own accounts and databases.
How long until something runs?
One automated step on real data comes first, before the full pipeline, so you see results early and decide whether to continue.
Which task does your team repeat every day?
CS-01Describe the process, the tools involved and what a mistake costs.
- What the product does and who uses it.
- What exists already: designs, code, a backend, a store account.
- Your deadline, and anything fixed such as a launch date or a platform.