What inspired us

The gap between learning and earning in tech is broken by a paradox: employers want experience, but you need a job to get experience. Course platforms give theory without stakes; freelance marketplaces give stakes without supervision. We wanted a third path — a studio that hands students real, supervised, paid project work and gives employers checkable proof of what was delivered.

What it does

PraxisAI is an AI-operated apprenticeship studio. Google Gemini runs the core knowledge-work of the studio — turning a raw client brief into a bounded scope, decomposing it into a delivery plan, and performing QA (including multimodal QA on screenshots and PDFs). Humans approve every consequential decision. Completed work produces a cryptographically signed, publicly verifiable credential.

How we built it

A modular monolith: a Next.js web app, a FastAPI transactional API over 100 PostgreSQL tables, and a typed Gemini adapter (google-genai) that enforces structured output via response_schema so malformed model output fails closed instead of entering the database. Every agent run is recorded append-only with model, prompt version, token usage, latency, and a correlation ID. The full Google Cloud footprint (Cloud Run, KMS, Secret Manager, Storage, Scheduler, Monitoring) is defined in Terraform.

What we learned

How to give an LLM real operational responsibility safely — enforced schemas, versioned prompts, treating every brief as untrusted data, prompt-injection containment, and a strict human-approval boundary for anything touching money or credentials.

Challenges

Keeping it honest. We are pre-revenue and pre-launch, and we deliberately refused to invent users, revenue, or partnerships — every demo surface is labeled "Demo data," and the repo documents exactly what is built versus not. Building trustworthy AI operations was harder, and more valuable, than building more features.

Built With

Share this project:

Updates