GenAI · own product · 2026

PrimeroLab

A study app built around answering. Every answer, spoken or typed, is graded from 1 to 5, and what went worst comes back first.

Visit PrimeroLab How an answer is graded

Built with

Next.js 15
React 19
TypeScript
Supabase
Postgres
pgvector
Gemini
OpenAI
Claude API
Tailwind
Vitest
Claude Code

60 slevel diagnostic on the landing, no sign up

1–5score on every answer, with what was missing

2model providers, with fallback

224automated tests

The problem

Rereading feels like studying, but never shows what you could not explain. Flashcards help, but you write them and grade yourself.

What a study session looks like

Decks come from your material or a topic. Each answer is graded on the spot, with what was missing and a complete example.

A card is answered out loudor typed

The audio is transcribedWhisper

The answer is graded from 1 to 5what was right, what was missing, and an ideal answer

The score is checked and saveda malformed grade never reaches the schedule

The next review is scheduledwhat went worst comes back first

language modelcodeuser

Besides decks, there is a technical interview simulator with a crew of interviewer characters, and an express mode with a time limit for quick reviews.

Any material, not just the first pages

The first version cut every document at 6,000 characters: 95% of an 80-page PDF never reached the model. Now the whole document is chunked and searched with pgvector.

Chunks of about 1,200 characters with 150 of overlap, so an idea split at a boundary is not lost.
An HNSW index for similarity search, with a documented fallback when the extension version does not support it.
The owner's id sits on every chunk, so the security check needs no join.

A study assistant that reads your progress

The study chat used to stuff keyword matches into every prompt. It was rebuilt as a tool-calling agent: “what should I review today?” reads the real queue.

The user asks a questioninside a workspace or with a deck open

The route builds the tool context from the verified sessionuser id, workspace and deck

The model chooses toolssearch study material, list decks, cards due for review, study progress, search documents

Each tool queries Postgres with the user's own clientonly the user’s own rows

The model answers with what the tools returned

language modelcodeuser

Security lives in the database

Every tool runs with the user's own session, so row-level security decides what comes back. The model never supplies a user id.

Hard limits

4 tool rounds at most, so a runaway loop cannot run up the model bill.
8 seconds per tool and 6,000 characters per tool result, so one slow or huge answer cannot stall the conversation.

The assistant does not go down with one provider

Gemini answers first and OpenAI takes over if it fails. Every reply logs which one answered, so a silent fallback never hides.

Why what went worst comes back first

Each 1 to 5 score feeds SM-2. FSRS retains better, but it needs review history that does not exist yet. SM-2 builds it.

Built and run by one engineer: product, data model, AI features, security and deploys.