Context
trophi.ai is an AI coaching platform for sim racers. Before this project, every coaching insight lived in a post-session report. In interviews and on-rig sessions, drivers kept asking for something closer to a coach in the car with them, because reading another dashboard after a race was rarely something they actually did. Mansell AI was the response to that pattern.
Problem
Telemetry feedback was arriving too late to influence behaviour. By the time a driver opened a report, the muscle memory from that lap had already set. We needed a coaching layer that worked in real time, respected how little attention a driver actually has at race pace, and still had enough personality to belong on a streamer's broadcast instead of fading into the background like a generic system voice.
My role
I owned the voice coaching UX, the in-car HUD surface, and the branded identity work that made Mansell recognisable on stream. I worked closely with engineering and the data-science team on the underlying coaching model, mostly around how cues were prioritised and timed.
Goals
- G1Deliver coaching cues during the lap, so a driver can self-correct on the next corner instead of waiting for a debrief.
- G2Keep cognitive load low enough to use at race pace, since anything that pulls focus from the apex is unlikely to survive past a few sessions.
- G3Give Mansell a distinct, branded presence that earns its place on a streamer's broadcast, while still feeling like part of the driving experience instead of a marketing overlay.
Process
- 01Shadowed sim racers across skill levels in their own rigs to understand when in a lap a spoken cue would actually be heard instead of tuned out.
- 02Mapped the corner-by-corner attention budget so audio cues could land in the small windows between corners and on the cool-down lap.
- 03Prototyped voice and HUD coaching in parallel and compared cognitive load. Voice consistently held up better, which is what eventually pushed it to the primary surface.
- 04Worked through telemetry latency with engineering, because a cue that lands a corner late does more harm than no cue at all.
- 05Iterated on Mansell's tone, pacing, and visual identity with streamers in mind, so the coach reads as a character on broadcast instead of a system notification.
Key design decisions
Decision 01
Voice over HUD
On track, the driver's eyes belong on the apex. Voice coaching became the primary channel because it kept focus where it had to be while still delivering specific, actionable feedback.
Decision 02
One cue per corner
We capped Mansell to a single high-leverage prompt per turn. In testing, it made the coaching feel considered instead of chatty, and drivers started actually acting on the cues.
Decision 03
A character, not a system voice
Mansell was given a name, a tone, and a small visual identity on purpose. It had to feel like part of the racing experience for the driver and hold up on a stream, although the goal was personality, not branding for its own sake.
Turning point
Session Preferences: handing control back to the driver
Early versions of Mansell automatically generated coaching across a driver's weakest areas for the entire track. It was technically powerful, however in user interviews and on-rig usability testing it was consistently described as cognitively overwhelming. Drivers were getting too many corrections in parallel, so very little of it was actually being retained between laps.
What that feedback pointed at was simple enough: more AI output was not always better. Drivers wanted agency over what was analysed, which sections of the track were coached, how verbose Mansell was, and what coaching style they received, because deliberate practice on one or two skills tended to be more useful than a wall of generalised feedback.
1.Select a Reference Lap and coach mode to be trained on by Mansell:
2.Select the sections of the track on which you wish to receive feedback:
3.Select the techniques you want Mansell to focus on:
4.Select your desired level of verbosity from Mansell:
Feedback Points Per Section
“I can’t actually work on anything specific because Mansell is talking to me about five different things every lap.”
Session Preferences was introduced as a personalisation and control layer instead of a settings dump. The goal was to reduce information overload, make feedback genuinely actionable, and let drivers focus a session on the specific skills they were trying to build. In practice, that meant choosing a reference lap and a coach mode, picking the exact track sections to be coached on, narrowing the techniques Mansell would speak to, and dialling the verbosity up or down per session.
The redesign turned Mansell from an always-on commentator into a controllable coaching experience. It protected the driver's attention budget, supported long-term learning and retention, and made the coaching feel like deliberate practice instead of noise.
User flow
Step 1
Driver launches a session and Mansell calibrates to the car and track in the background.
Step 2
During the lap, spoken cues arrive in the headset corner by corner, paced to the natural rhythm of the circuit.
Step 3
After the session, the full timeline of cues is replayable alongside the telemetry that triggered them.
Outcome / impact
“Man, Mansell literally taught me the track to Laguna Seca. I turned it on, went in cold, and it just told me exactly what I needed to know and what I should do…Because of trophi.ai I have the ability to learn.”
Reflection
Mansell shifted improvement from a post-session analytical exercise into something closer to in-the-moment learning. For a lot of users, it is also the specific reason they sign up for trophi.ai in the first place, which is why it ended up driving most of acquisition, conversion, and retention, and contributed materially to the product passing USD $1.5M in ARR in its first year in sim racing. The thing I took away from the build was how much restraint mattered, because in a domain where users are genuine experts, every cue we cut made the coach feel more competent.
Next project
Real-Time Skill Assessment