Why it was needed
The problem
Companies spend a fortune on training, and most of it evaporates within weeks. People leave the room with good intentions and go back to the same habits.
Dr. Pelè had the method and a book behind it. He needed software whose job is to change how people behave at work, and a product he could sell to organisations.
The system
What I built
An AI coach built on the author's method
After a program ends, the coach keeps working with each person for weeks. It is grounded in Dr. Pelè's own material, so the guidance follows his method instead of generic advice.
Memory that holds a long relationship
Every turn, the coach assembles its context from a rolling summary, the most relevant past messages found by vector search, the recent conversation and the commitments being tracked. Prompts stay small and fast however long the coaching runs.
Dashboards that tell leaders the truth
Conversations become structured priorities, behaviours and results that flow to leader and executive dashboards. When the data is thin, the dashboard says "not enough signal yet" instead of inventing a number.
Privacy and tenancy from day one
Coaching data is de-identified before it reaches any dashboard, and row-level security keeps every organization's data separate. Five roles, from participant to executive, each see only what they should.
Step by step
How it runs
A program ends
An organization runs the program with a cohort. Each participant leaves with commitments they made in the room.
The coach follows up
Over the following weeks the AI coach checks in with each person, asks what they did, and works through what got in the way.
Conversations become data
Background jobs extract priorities, behaviours and results from each conversation into structured records.
Leaders see the pattern
De-identified results reach the leader and executive dashboards. Where the data is thin, the dashboard says so.
Where the judgment went
The hard parts
Long memory, small prompts
A coaching relationship runs for weeks. Sending the whole history to the model on every turn gets slower and more expensive the longer someone stays. Retrieval scores past messages on recency, similarity and the commitments they mention, and older turns roll into a summary.
An honest dashboard
A leader who sees a confident number acts on it. With three conversations behind it, that number is noise. The dashboard shows nothing before it shows something false, and groups too small to stay anonymous are never shown on their own.
What came of it
The outcome
10 weeksfrom concept to launch
- Concept to launch in 10 weeks.
- Live cohorts running inside enterprise organizations today.
- A product the author sells to organisations, built on his own idea and methods.