ProjectsB5
Evolutia
Adaptive training where the AI proposes and the person decides.
- Role
- Creator / Product Lead — AI-assisted development
- Since
- July 2026
- Status
- In development, with a private Android demo.
SummaryAerial view
Evolutia plans, logs and adapts strength training. An AI suggests changes to the next week, and nothing happens until the person reviews them. It started at the end of July 2026, when I started training in a conventional gym after CrossFit.
Review an AI proposal
Synthetic dataThe AI proposes three changes for next week. Accept, edit or reject each one, then create the draft.
This demo needs JavaScript. The case below reads the same without it.
CaseOn foot
Problem
Logging and adapting training is only useful if an automatic suggestion cannot break the rules of the plan or change a week without anyone looking at it.
Decisions
The AI proposes, the person applies
BusinessA suggestion is a document to review, not an action. The person can accept, edit or reject each change.
TechnicalThe model generates a TrainingProposal. Applying it creates a new draft week. It is never published or activated automatically.
Deterministic rules around the model
BusinessThe plan's rules hold whatever the model suggests. The AI works within them.
TechnicalRules and permissions live in Postgres with row-level security and RPC functions. Edge Functions call OpenAI and validate what comes back.
Test what cannot be checked by eye
BusinessBefore showing it to anyone, the planning, logging and proposal flows had to pass automated checks.
TechnicalThree separate suites: Bun for logic, pgTAP for the database and Playwright for end-to-end flows.
How I checked it
The planning, logging and proposal flows passed three automated test suites, and there is a private Android demo.
What I learned
Putting a person between the proposal and the plan made the product simpler to reason about. The rule that an apply only creates a draft removed a whole class of problems.
Role
Product, plan rules, limits of the AI, requirements, acceptance criteria and review. Codex implements a substantial part, with cross-review between ChatGPT and Codex.
EvidenceBelow the surface
Architecture
- Data
- Person
- AI model
- Storage
The same flow, step by step
- The plan and the logged sessions are stored under database rules.
- OpenAI receives the context and returns a TrainingProposal.
- The proposal is validated against the plan's deterministic rules.
- The person reviews, edits or rejects each change.
- Applying creates a draft week. Publishing it is a separate, human step.
Product plates
Plate 1 of 6Opening screenEditorial compositionSynthetic data Plate 2 of 6The weekEditorial compositionSynthetic data Plate 3 of 6Logging a sessionEditorial compositionSynthetic data Plate 4 of 6ProgressEditorial compositionSynthetic data Plate 5 of 6Check-inEditorial compositionSynthetic data Plate 6 of 6Proposal, review and applyEditorial compositionSynthetic data
Stack
Page last revised: