Internal guide. Enter the password to continue.
We point the system at a brand. It gathers what the public says about it — social media, reviews, news, search, filings — measures it, and writes a report for investors. Every claim in the report links back to the source it came from. A full run takes about an hour.
Someone fills in a short form: the brand name, where it operates, any competitors they already know about, and — most usefully — what the investor actually wants to understand.
Why: knowing the real question up front keeps the report focused instead of generic.
The system does a fast, cheap scan to confirm it has the right brand, find the real competitors and the parts of the market that matter, and sketch a plan for the main research. It does not draw any conclusions yet.
Why: this takes a minute or two. It stops the slow, expensive stage from running on the wrong brand or the wrong competitor set.
Before the heavy work begins, the system pauses twice and asks the user to confirm: first the list of competitors, then the overall scope. Once both are confirmed, it runs on its own.
Why: fixing the scope here takes a minute. Fixing it after a full run wastes an hour. These are the only two points where a person is involved.
Gather. All at the same time, it pulls from every source it has:
Measure. Every comment is sorted by how people feel, what it's about, and what they want (a fast, low-cost model does this). The actual counting and charting is then done with code — not guessed by the AI — so the numbers are real.
Write. Only now does the system form a view and write the report, putting a link to the original source next to every claim.
Why: gather widely first, count with code, and form the opinion from the data — the way a good analyst works.
A separate pass reviews the draft like a fresh reader, looking for weak evidence or missing points, and corrects them. It then tidies up, writes the summary, adds a cover image, saves screenshots of the key sources, and locks a final version.
Why: the model that wrote the report is not good at judging its own work. A fresh review catches mistakes before the report goes out.
The finished report comes in three forms: a web page with clickable sources and live charts, a PowerPoint for meetings, and a zip of all the raw data. Users can ask for changes in plain English, and every version is saved so nothing is ever lost.
| Job | Model | Why this one |
|---|---|---|
| Probe | Claude Sonnet 4.6 | Quick, capable scoping and planning. |
| Analyse | Claude Sonnet 4.6 | Runs the sources, reads the evidence, writes the report. |
| Sorting comments | Claude Haiku 4.5 | A lot of repetitive labelling — a fast, low-cost model fits. |
| Deep research | Perplexity · Gemini · OpenAI o3 · Grok | Four tools at once; each finds sources the others miss. |
| Revise | Claude Sonnet 4.6 | Reviews the draft, then polishes and finalises. |