How I Built Supercomp.app, an Agentic Fitness Planner (in 3 Months)
Why This Exists
Supercomp wasn’t born out of a startup pitch deck or a market gap spreadsheet — it came from burnout.
I was working 14‑hour days, allocating zero time to my health, and slowly spiraling. The irony was obvious: I knew fitness works when you follow a plan, but I didn’t have the time, money, or mental bandwidth to create or manage one properly.
I tried:
Fitness coaches → great, but $150+/month wasn’t realistic
Macro trackers (e.g. Cal AI) → good at logging, bad at guiding
What I needed didn’t exist: a system that told me exactly what to do, every day, end‑to‑end — training, meals, macros, cardio, expectations — at a reasonable cost.
So I built it.
In ~3 months, I went from 80kg → 72kg, kept my strength, preserved my frame, and dropped visible fat. That was enough proof for me that the system worked.
This post is about how I built Supercomp.app technically, using modern AI agent principles — not hype, not buzzwords — but real architecture decisions, trade‑offs, and lessons.
High‑Level Architecture
At a high level, Supercomp is not a single prompt → single response app.
It’s a multi‑agent system where each agent has:
A clear responsibility
Deterministic constraints
A shared understanding of user goals
Think less ChatGPT wrapper, more coordinated planning system.
Core Components
Frontend: React / Next.js
Backend: Node.js (API‑first)
AI Runtime: Groq
Models: LLaMA‑based reasoning models (via Groq)
Memory / Context: Lightweight RAG + structured state
Data Layer: Postgres + vector store
The Core Insight: Fitness Is a Planning Problem
Most fitness apps treat fitness as a tracking problem.
Supercomp treats fitness as a planning + execution problem.
That single distinction drives the entire architecture.
You don’t need AI to count calories. You need AI to reason about trade‑offs over time:
How aggressive can the deficit be without muscle loss?
When does cardio interfere with recovery?
What happens if a user misses a day?
How do weekly macros stay coherent across meals?
This is where agents shine.
Agent Design
Instead of one large prompt, Supercomp uses specialized agents.
1. Goal & Physiology Agent
Responsible for:
Interpreting user stats (weight, BF%, height, activity)
Defining realistic transformation targets
Setting non‑negotiable constraints
Example outputs:
Target rate of fat loss
Protein floors
Maximum cardio volume
This agent never generates meals or workouts. It defines the sandbox.
2. Nutrition Planning Agent
Responsible for:
Generating daily meal plans
Hitting macros within strict tolerance
Respecting food preferences
Key rule:
The agent is evaluated against the same targets it was given
This avoids what I call numeric truth drift — where generation and validation disagree.
Meals are generated deterministically where possible, stochastic only where safe.
3. Training Agent
Responsible for:
Weekly split design
Volume & intensity allocation
Progression logic
The agent understands:
Recovery debt
Interference from cardio
User experience level
It does not adapt daily yet — that’s v2.
4. Cardio Agent
Cardio is handled separately because it’s the easiest way to break plans.
This agent:
Calculates burn via MET‑based formulas
Keeps cardio predictable
Feeds calorie offsets back into nutrition
Once the template is fixed, burn is deterministic.
Why Groq
Groq wasn’t a hype decision — it was a latency decision.
Planning requires iteration:
Generate
Validate
Adjust
High latency kills this loop.
Groq allowed me to:
Chain agent calls
Enforce retries
Run validations inline
Without turning the UX into a loading screen simulator.
RAG (But Not the Buzzword Version)
Supercomp uses RAG sparingly.
What’s retrieved:
Exercise definitions
Nutrition references
Prior user plans (structure, not content)
What’s not retrieved:
- Raw user logs dumped into context
Agents operate on structured state, not chat history.
This keeps reasoning clean and cheap.
Guardrails & Validation
Every plan goes through:
Macro validation
Calorie tolerance checks
Weekly coherence checks
Recovery sanity checks
If validation fails:
The agent retries
Targets are re‑fed explicitly
No silent degradation
Failure is loud by design.
Why This Isn’t a Coach Replacement
Supercomp doesn’t replace great coaches.
It replaces:
Guesswork
Inconsistency
Fragmented tools
It’s for people who:
Are busy
Need structure
Will execute if told exactly what to do
Results & Proof
In 3 months:
80kg → 72kg
No strength loss
Visible fat reduction
The system worked before it was a product.
That’s the only reason it exists.
What’s Next
Adaptive plans
Missed‑day recovery logic
Deeper personalization
But the core philosophy won’t change:
If you know exactly what to do, every day, you will get results.
Final Thoughts
Supercomp is priced at $10 one‑time because this shouldn’t be gated.
If it helps even a fraction of people escape the spiral I was in, it’s worth it.
👉 https://supercomp.app
Thanks for reading.