In regular use
One research engine, many jobs
A question goes in. Four AI models research it without seeing each other's work, and every fact that matters is checked against its source before I act on it.
- 63
- key facts checked on the car lease research
- 5
- of them wrong, caught before I relied on them
- 4
- AI models researching blind to each other
Most research I need has the same shape, whatever it is about: find out what is true, from sources I can check, before I spend money or someone's time on it. I used to do that by asking several AI tools the same question and reconciling their answers by hand. The reconciling was where mistakes got through. In the run that started this, two of three tools gave the same figure to the wrong person. The majority was wrong.
deepreach does that job the same way every time. The layers that check each fact, and what each one has caught, have their own write-up. This page is what it gets used for.
Leasing a car
The fullest example is personal. In September I used it to choose and lease a new electric car. I gave it a rough question, it turned that into a research brief, and I edited the brief before anything ran. Four models then researched the cars, lease prices and owner reviews without seeing each other's work. Then the step that matters: 63 key facts, from prices to specifications to company details, were checked against their sources. 48 held up fully, 8 partly, 2 could not be confirmed, and 5 were wrong.
The wrong ones are the reason the checking exists. One model returned prices, a company registration number, review counts and energy tariff rates, every one marked "confirmed" and every one wrong. The real prices sat on listing pages that only appear once a browser runs them, and only an agent reading those pages in a real browser could see them. Even then, the top pick's price had changed when it was re-checked the same evening. Lease deals can disappear within hours, so every price now carries the date it was seen.
Once I had chosen, the same tooling carried the paperwork. It read the broker's emails, listed what had been answered and what had not, and drafted my replies; I edited and sent them. It pre-filled the finance application in my browser from my own records, one section at a time, with me confirming each. I entered the bank details, answered the declarations and signed. Before I signed the order, it checked every term on the order form against the agreed quote. They matched.
The same method, other jobs
- Guides with prices in them. The sauna directory I run publishes buying and business guides, including what a sauna costs, booking systems for sauna businesses and mobile versus fixed-site businesses. A price or a regulation reaches a reader only after a live check.
- Fund work for my day job. Maps of property funds and the firms that manage and value them, for business development at my employer. One started from 205 fund managers and settled which funds were open- or closed-ended from national fund registers, leaving 117 in scope. One covered every listed fund in a single Asia-Pacific sector, 11 in all, from about 48 primary filings, working out when each valuer appointment is due to rotate under the governing code. One scored and tiered 154 firms across Europe, the Middle East and Africa, with the valuer-rotation rules for each country.
- Other market maps. Lists of firms and people for the same work. One covered 52 owners and investors in a single sector across Asia-Pacific, with every name graded by how well it was confirmed.
- A senior hire search in Germany, for my employer.
- Smaller decisions, such as a computer purchase, whenever being wrong would cost money.
Where it stops
It researches and drafts. It does not send, sign or pay: an email, an application or an order goes out from me, after I have read it. And it does not make the decision. It hands me checked facts with a grade on each, and the choice is mine.