Fulkram
Fulkram / problems I fix / buried knowledge
One of the problems I hear most from owner-run businesses

"The answer's in there somewhere."

In a call recording, a doc, a thread from last March. Somebody knows it exists, and somebody has to go and dig it out by hand. I had the same problem at a fairly silly scale, so I fixed my own first. The receipts are below.

The expensive part is a person. The person who knows becomes the search engine: every question interrupts them, every answer gets given twice, and when they're away the answer's just gone. Meanwhile the recording that settles it already exists. Nobody's going to scrub through an hour of video to find one minute. So the knowledge sits there, paid for and unused.

The receipts

My own archive, indexed and put to work

Everything below ran on my own machine. The recordings never left it.
The knowledge base dashboard: 5,031 transcripts, 36.8 million words indexed, and the pain-point distribution with labels withheld.
The knowledge base dashboard. 5,031 transcripts from 7,801 recordings, indexed and searchable. SCREENSHOT from the live system.

Words, indexed and searchable to the second

The whole archive transcribed into 241,371 searchable chunks and 4,867,242 timestamped segments. Ask a question and get back the recording and the second it was answered. Transcription and indexing both ran on my own machine, so nothing sensitive went near anyone's cloud.

Where it livesthe case study

A wrong answer maps to the right lesson, to the second

On my exam platform, every wrong answer is auto-mapped to the lesson that teaches it and the exact second it's taught, snippet attached, flagged for a human to check. That's the job an indexed archive can do once it's searchable to the second.

Where it livesthe exam platform

It found what nobody had found by hand

Mining the archive for customer language surfaced pain points, in customers' exact words, that years of being in the room hadn't caught. Those phrases now do the marketing.

StateSCREENSHOT the pain-point list

Recordings became articles you can read right now

Posts live on my own site, every one generated from a recording, published and picked up by search. Go and have a look.

Open it nowLIVE dylanayaloo.com/blog
  • The extraction demo, the same shape on business documents: it answers with source chips, and where the sources disagree it flags the disagreement. Invented data, and it says so on its face. PROTOTYPE /demos/extract/

What I'd tell you not to build

The chatbot, first. "Train an AI on our stuff" usually turns into a chat window nobody uses, sitting in front of an index nobody built. The index is the asset. Get the corpus transcribed, chunked and searchable first, and in roughly half the cases I've seen, that alone settles it, because what people wanted all along was an answer with a source and a timestamp. Sometimes the right buy past that point is nothing.

And to be straight about scale: nothing here has served thousands of people at once. This corpus serves one business, mine. If you need ten thousand concurrent users, I can't show you a receipt for that today, and we'd need a proper architecture conversation before I promised you anything. I'd rather tell you that before you've paid me a pound.

Your team runs it without me: answers show their sources so a person can check them, and it's handed over documented.

First piece is fixed price, typically £750 to £1,500. Then we extend or stop. How pricing works →

Tell me where your answers are buried. Get in touch →