I wanted to know if a machine could become the smartest business mind I know. So I gave one thousands of hours of business teaching, and a perfect memory for where every word came from. Ask it a hard question and it does not guess. It finds what the experts actually said, and it shows me where they said it.
It runs all day on its own server. My other AI agents ask it questions, the same way I do. It has been torn down and rebuilt until it could win a blind test. The version running now beat the first version 63 to 11 with the labels hidden.
What it is
I had thousands of hours of business teaching sitting in transcripts, and no good way to ask it questions. A normal AI answers from memory and guesses when it should not. So I built Thoth on a method called GraphRAG, where the AI reads every transcript once and builds a map of everything inside. Every person, method, and number, and how they all connect. The map holds 101,556 entities and 73,187 relationships. In plain terms, over a hundred thousand distinct things it found, wired together by how they relate. When I ask a question, it walks that map and hands back the exact words the expert said, with the source attached.
The library underneath is 4,639 transcripts, about 30 million tokens. Tokens are the units AI reads, roughly three quarters of a word each, so that is hundreds of novels' worth of teaching. The software is my own maintained fork of Microsoft's GraphRAG, taught to look for business concepts instead of the news-style names and places it ships with.
How I prove it
Anyone can claim their system works. I make mine win blind tests. The same 40 questions go to two versions of the system, every answer is written down with no name attached, and judges pick winners without knowing which version wrote what. It is a taste test with the labels sealed until the votes are in. The current Thoth beat the first version 63 to 11 under those rules, winning every category of question. The judges are AI models too, from two different families, because one judge is an opinion and two is a measurement.
The numbers get a harder check. All 27,469 metrics the graph extracted are verified by a script against the transcripts they came from, and anything that does not match gets flagged for human review. That script needed checking too. Its first version flagged three times too many, because spoken numbers like a million five confused it. Validators need validators.
Any AI · from memory
There are many great ways to attract customers. Popular options include social media marketing, networking events, paid advertising, referral programs, and content marketing. The best mix depends on your goals and budget.
Thoth · from the corpus
There are only four. Warm outreach, content, cold outreach, and paid ads. A two by two: do they know you, and do you reach one or many. “They cannot buy it unless they know you exist.”
Source 2782Entities 66737 · 42606
Until a million a year, run one channel. If you run two, make it content plus warm outreach.
Source 2783
The right side is a real Thoth answer, trimmed for length. The source numbers are its own citations, and each one reopens a real transcript.
The rebuild that lost
The best story in this project is a failure. The second version of Thoth was the disciplined one, with stricter rules and cleaner structure, and I expected it to win easily. It lost its blind test to the messy first version, 8 to 2. Instead of guessing why, I did forensics. The corpus had quietly failed to download most of three source catalogs. And the extraction prompt literally ordered the AI to stay sparse, a few concepts per chunk and no more. Run on identical text, the old messy version had found six times the entities and twenty times the connections, and when I measured its supposed noise, it was under three percent junk. The big messy graph was almost all real.
The fixes went cheapest first. Two small changes, giving retrieval a bigger share of the model's reading room and porting over the better answer prompt, flipped the blind result from losing to winning in a single commit. Only then came the expensive fix, a full rebuild, and even that was proven on a small slice of documents before the real money was spent. The lesson stuck. When a measurement surprises you, it is trying to tell you something.
Every seat earned blind
Thoth runs on several AI models, each holding a different seat. One writes the answers, one summarizes, one judges. No seat was assigned by reputation. Each was won in a sealed bake-off, with answers written to disk and verdicts recorded before the labels were opened. The answer seat took 63 head to head comparisons across 8 candidate models, and the winner later defended its seat 12 to 0 in a rematch.
Just as often, the test said no. A fancier configuration tied its blind test at 1.67 times the cost, so the cheaper one shipped. Three separate results saying more compute is not better landed in a single day, and none of them shipped. The test decides, not the price tag.
What a question costs
Thoth runs around the clock on its own server. A fully answered question costs about ten cents and comes back in about two and a half seconds. Adding new transcripts costs about nine cents an update.
The cheap answers sit on a heavy foundation. Building Thoth consumed over 680 million tokens across more than 365,000 model requests, against a library of about 30 million. Between extraction, rebuilds, and testing, the library was effectively read more than twenty times over, so that every question after it costs a dime.
Questions about this system, or the problem it could solve for you?
Discuss your project