Real-Time AI

From client's words to a live map, so the coach can stay present.

Status: working prototype, validated in coach testing. Extraction, mapping, and coach override are live. The learning loop is in active testing.

Hero visual pendingThe live canvas: client phrases mapped across the four quadrants mid-session. The product in one image.

Placements accepted: 88%Placements accepted
Would use it in real sessions: 92%Would use it in real sessions

See how we measured this

Context

In a live session, a coach has to make a bad trade: take notes and miss the moment, or stay present and lose the pattern. The moments that slip past are the ones that matter most. The instant a client moves from struggle language to solution language is the whole opportunity in solutions-focused coaching, and it is exactly the moment note-taking causes a coach to miss.

The foundation is the Dialogic Orientation Quadrant, developed by Haesun Moon (opens in a new tab): struggles below the line, strengths above, past on the left, preferred future on the right. Her framework maps the conversation. My contribution was making it read itself in real time.

My role: lead AI product designer, and the person who built it. I owned the system architecture, UX, behavioral design, and the NLP pipeline.

The real question wasn't how to transcribe a session better. It was how to make what matters visible in real time, readable in a glance, so the coach's attention stays with the client.

The quadrant model: the Dialogic Orientation Quadrant, adapted for real-time extraction.

Objective

Make the meaningful parts of a conversation visible as they happen, without costing the coach their attention. The coach needed three things:

Capture without cost: a way to record meaningful client language without breaking eye contact or presence.
Trust in what they saw: classifications they could check and override, not a black box telling them what a client meant.
A glance, not a study: something they could orient to in about two seconds before returning to the person in front of them.

All of it had to hold up in a high-risk setting. Classifying a person's words in a live, quasi-therapeutic conversation means a confident wrong call doesn't just annoy the user. It breaks the trust the whole tool depends on.

Approach

1. The system proposes, the coach decides

The system reads the session transcript as it's generated, extracts client phrases, and places them on the quadrant. The coach sees every placement and can move any phrase to a different quadrant with a tap. This is the constraint that makes an AI tool usable in a human moment.

Phrases reach the map about 6 seconds after they're spoken early in a session. That lag grows as the conversation gets longer and varies with the language model used, which is part of why the map is designed for occasional glances, not constant watching.

A phrase with its confidence score and override control.

2. Every placement can be traced

Each extracted phrase carries a visible confidence score and a classification path the coach can follow. No black-box summaries. A coach can always see why something surfaced, and reject it if it doesn't fit.

3. Only confident reads reach the map

Phrases appear on the canvas only when they clear a confidence threshold, set at 70% and adjustable. Everything else stays in a list for review instead of cluttering the map with guesses. That is what keeps a real-time map both trustworthy and readable at a glance.

The shift, visualized: a phrase moving from below the line (struggle) to above it (solution). The exact moment the tool exists to catch.

4. The system stores meaning, not transcripts

What persists is what mattered: extracted phrases, their placements, confidence scores, and how often a phrase recurs across sessions. Not a wall of text. That persistence is what turns a session aid into structured conversational state.

5. Every correction teaches the system

Override isn't only a safety rail. When a coach moves a phrase, or highlights something in the transcript the system missed, that correction becomes training signal. The coach staying in charge is the same act that improves the model.

This part is designed and partially built. I'm testing different approaches to see which one improves classification without eroding the coach's trust in the map. I'd rather test which direction is right than ship a guess.

6. A quiet signal shows where the conversation is heading

The canvas shows where phrases landed. The next layer shows where the conversation is going. As a session develops, the system recalculates the conversation's center of gravity and shows it as a soft heat zone on the quadrant, so a coach can tell at a glance whether the dialogue is drifting from struggle toward preferred future.

It is deliberately quiet, a single soft gradient rather than a busy heat map, because ambient awareness only helps if it doesn't pull the coach off the client. And it sits on top of the phrase data rather than replacing it: the individual phrases stay the source of truth.

This layer is designed and next on the build list.

Outcomes & Impact

In coach testing, coaches stayed fully present instead of choosing between note-taking and attention.
Coaches accepted 88% of placements as they were and corrected 12%, so the override was used when it mattered without becoming constant work.
23 of 25 coaches (92%) said they would use it in real sessions.
Override made the tool trustworthy enough to use live, with a real client.
The confidence threshold kept the map readable in about two seconds, with low-confidence reads still available for review.
Conversational state persists between sessions and is structured for downstream AI, not locked in a transcript.

Placements accepted: 88%Placements acceptedShare of system placements coaches left unchanged, across testing with 25 coaches
Would use it in real sessions: 92%Would use it in real sessions23 of 25 coaches in testing

Reflection

The real product wasn't transcription. It was structured meaning with a human in charge.

A transcript records what was said. Conversational state captures what mattered, in what context, and how often. Common Project is the layer that converts one into the other without taking the decision away from the human.

Speed is the next design problem.

A 6-second delay works for glances, but it grows over a long session. The next round of work is keeping the delay flat, by limiting how much of the conversation the model rereads each time and choosing models for speed as well as accuracy.

What started as a coaching aid was really infrastructure.

A system that understands what was said and what it meant, with a record of every decision and a human override, is exactly what downstream AI needs in order to be trusted. The coaching canvas was the first surface. The structured layer underneath is the reusable part.