Signal & Flow
An AI-powered digital strategy platform: multi-agent workflows that produce competitor analysis, UX evaluation and behavioural insight, using synthetic customer personas in place of a traditional research cycle. It's not a one-off build; the agents and prompts are continuously refined as real reports come back from real use. The persona layer is, in effect, a context system in miniature: curated knowledge standing in for research, shaping what each agent produces.
It's live and in production, not a concept, running on subscription tiers, with a pay-per-report option for one-off use.
Visit Signal & Flow →A recorded preview of the live homepage.
Why I built it
Competitor teardown, UX audits and customer insight work still mostly run on a human analyst's time, one engagement at a time. Signal & Flow is my attempt to see how far a multi-agent system can go in compressing that work, without hollowing out the judgement that makes it useful in the first place.
How it works
Under the hood, Signal & Flow runs a set of specialised agents rather than a single model call: one agent pulls apart a competitor's site and positioning, another runs a structured UX evaluation against it, and a third generates behavioural insight by simulating how different synthetic customer personas would actually experience the journey. Each agent is scoped narrowly on purpose: the output is stitched together into a single report rather than asking one generalist prompt to do all of it at once.
The persona layer is where most of the iteration happens. Getting synthetic personas to produce insight that's genuinely useful, rather than generic or overconfident, is largely a prompt engineering problem, and it's the part of the system I'm still actively refining. Underneath the prompts, it's really a context curation problem: what goes into a persona, what background, priorities and constraints each agent is given to work from, matters more than the wording of the prompt or the model call itself. Get that input wrong and no amount of prompt polish fixes the output.
My role
I designed the agent architecture, wrote and continue to rewrite the prompts that drive each stage, and built the product around it, including pricing, tiers, delivery. This isn't a proof of concept I built once and shelved; it's the project I use to stay hands-on with agentic AI systems, rather than only advising on them from the outside. This is context design with the stakes made concrete: get the input wrong and the output is immediately, visibly wrong, with no client meeting to soften the gap.
Where it's headed
Signal & Flow is live and in active use, and it's staying that way: new report types, persona refinements and workflow changes ship as I learn what's actually useful versus what's just interesting to build.
If a problem like this sounds familiar, I'd be glad to talk through how it might apply to your business.
Start a conversation