Case study · Apex Performance Labs · Colombo
The AI system we built for Apex Performance Labs.
A performance and longevity gym in Colombo. Our founding client. We built the operating layer: four AI agents on one shared member record.
The brief
A gym whose software had to match its floor.
Performance training, longevity work, clinical depth — and the same software stack every gym buys. Coaching stopped at the door. Labs landed as PDFs.
Coaching between sessions
Members train three hours a week and live the other 165. Coaching had to follow them home.
Labs acted on
Bloodwork and body scans had to reach the plan — read, flagged, signed off.
Leads answered in minutes
Leads go cold fast. Follow-up had to happen while the interest was real.
One view of every member
Plan, bloods, sessions, nutrition, payments, conversations. One record, every role.
What we built
Four AI agents on one member record.
Each owns a job. Each escalates to a human. Tap one.
the core
One member record
plan · bloods · sessions · life
the aim
Retention ↑
Performance ↑
Lena
the coach
Growth
fills the gym
Clinical
reads labs
Ops
the business
Lena agent
your coach · live
the coach
Lena
Coaches every member off their own data — plan, bloods, sessions, life.
See the workflow & integrations →
Growth agent
9 follow-ups drafted today
fills the gym
Growth
Turns leads into members without the chasing.
See the workflow & integrations →
Clinical agent
AI drafts · clinician confirms
reads labs
Clinical
Reads the panel, flags what matters — a clinician confirms.
See the workflow & integrations →
Ops agent
today
the business
Ops
Runs the floor and the admin, catches churn, and shows you the revenue.
See the workflow & integrations →
The connective tissue
One shared member record — plan, bloods, sessions, nutrition, payments, conversations. The agents hand work to each other. Staff see the whole person.
A closer look
Under the hood.
Every mechanism below is shipped and running.
The member app — web and native
Web and native Android. Push that opens the right screen; health data via Google Health Connect. QR check-in, live-session screen, booking, coach messages, VAT invoices.
Lena, up close
Logs a meal from a photo, a sentence, or “the same as Tuesday”. Counts the cooking oil nobody mentions. Says “Logged” only when the write happened. Clinical goes to a clinician.
The engine underneath
BMR by body composition. Deficits from rate of loss. Floors that refuse unsafe plans. An LLM composes; a deterministic validator checks; a clinician approves every plan.
Run a gym or studio? See custom AI agents for gyms and studios. Run a clinic? Start with custom AI systems for longevity clinics.
How it was built
Built on a live gym floor, feature by feature.
The model is forward-deployed: an engineer deployed inside the operation, AI carrying the heavy engineering. Build start May 2026; full stack live by July.
Forward-deployed
Engineers embedded in the operation — on the floor, in the schedule, in the numbers.
Shipped feature by feature
Small releases straight onto a live operation. What didn’t survive the floor was rebuilt.
Tested like software
Every Lena release gated on evaluation suites, including regressions built from real conversations that went wrong.
Run by its builders
The team that built the system still runs it in production, daily.
One rule held everywhere: AI drafts; humans approve. Nothing clinical ships without a clinician signing off.
Modern web stack · Gemini models on Google Vertex AI · enterprise cloud hosted in the EU/UK · WhatsApp, SMS, email and push rails · GDPR and Sri Lanka PDPA-aligned data practices.
Learnings
What building on a live floor taught us.
Six things Apex taught us. Each is now a rule in how we build.
The confirmation gate beat the model upgrade
“Logged” only ever means logged.
LLMs don’t do arithmetic here
Every number a member sees is computed deterministically from their own data. The model presents it.
Ship onto a live floor, in small pieces
Features that don’t survive real staff and members get found out in days.
Localise or it doesn’t get eaten
Meal plans landed once built from national food-composition tables, local dishes and each member’s allergies.
Test AI like software
Releases gate on regression evals built from real conversations that once went wrong.
Humans-in-charge is architecture
Clinician sign-off shaped the system: review queues, approval gates, escalation paths.
Honest numbers
The numbers we’ll put in writing.
All real, all current. No dressed-up percentages.
~3
months, build to production
Build start May 2026; full stack running the operation by July. A scope conventionally quoted at 18–24 months.
88 of 88
active members coached
Every active member at Apex is coached by Lena. The whole floor.
13,651
coach messages
Coaching messages sent by Lena — and counting.
Live
in production, end to end
Member app, operator console, clinical pipeline, growth engine — running the operation daily.
Across the whole member · live counts
1,179
meals logged — 555 from photos
179
body-composition scans tracked
141
weight readings
413
sessions booked
Nutrition, scale readings, training — one coach, reading one record.
What it sounds like · real coach messages, names removed
“Your labs are in: all markers are in range. Tap to see the full panel.”
“Your new scan is in — skeletal muscle up 0.4kg, fat mass down 0.7kg. A solid week of effort showing up exactly where it counts.”
“Bloodwork looks reassuring and your current plan is appropriate — the dizziness was most likely mild dehydration. Keep your water up around your sessions.”
“Your latest scan is in — a steady week, completely normal as bodies move in waves. Your body composition held steady, which means what you’re doing is working.”
Counts from the live production database, 23 July 2026. Engagement signals, not outcome proof. We publish no revenue or retention claims.
The vision
Insight-first — where all of it is heading.
Everything above is live. Insight-first is where the practice is heading — these surfaces are in build at Apex, validating before they ship:
Insight-first home
The home screen leads with what changed against the member’s own history — protein pace, weight trend with a goal ETA, training rhythm. Deterministic detectors compute every number; the AI never invents one.
Promises, kept
Lena makes no promise she can’t keep. A commitments rail turns “I’ll check in tomorrow” into a scheduled, delivered follow-up.
Action, through guarded doors
Lena is gaining the ability to book a member’s session herself — through the app’s single, guarded booking door.
Full disclosure
Apex is our founding client.
Stead AI grew out of building this stack. Our founder built Apex’s system first; the firm formed around what that build proved. We still run it in production, every day.
Every claim on this page is checkable against a live operation. More on the company and the founder: about Stead AI.
Start a conversation
Your operation next.
A system like this, designed around your operation.