Case study
Proficio
Proficio turns a two-minute intake form into a personalized, day-by-day study plan for Brazil’s ENAMED — and ten more medical-residency exams. A deterministic engine decides where every study hour earns the most, grounded in real question-incidence data from eighteen ENAMED papers (2011–2025); the writing is generated to match. A joint effort, live at useproficio.com.
How it works
Forecast-driven hours
Each study hour is allocated where it earns the most expected score, weighing how often a topic appears, how confident you are, and real coverage data.
Focus on what’s tested
The seven official areas are weighted by real exam frequency, and within each, the dominant subtopics come first — the top few themes per area cover the majority of the exam.
A schedule you can trust
Weeks, daily slots, and the order of study are pure, reproducible math — no randomness, and every number traces to a source.
Grounded writing, never invented
The guidance is generated against a strict schema with a critic review pass, and falls back to a deterministic outline rather than make anything up.
Under the hood
Plain English by default — switch to the engineering detail.
A deterministic scaffold
The plan’s structure — weeks, hours, sequencing, mock-exam placement — is computed entirely by math, with no AI in the loop. It’s instant, free, and reproducible.
A Python “brain” ingests sourced canonical data (areas, themes, references, rules) and applies Rasch-style scoring: per-area priority is exam_weight × (6 − self-rating) + a floor, distributed by normalized priority. No Date.now or Math.random; determinism is held by golden tests.
Hybrid generation
AI writes only the prose — area copy and daily guidance — inside the locked structure; it never rewrites the strategy.
One Anthropic SDK call per area, each schema-validated with Zod and checked by a bounded critic loop (≤3 passes); output is cached by input hash so re-runs cost nothing, and any failure falls back to the skeleton.
Reproducible by design
The same intake always produces a byte-identical plan — which matters for a product advising students on a high-stakes exam.
The render date is injected via env, output is input-hash-cached and versioned, and Zod contracts enforce every boundary between the Python brain and the Node request path.
Built to fail safe
An AI timeout or a rejected draft never breaks a student’s plan — the deterministic skeleton always renders.
On any model error, critic rejection, or parse failure the block is dropped and the scaffold renders; every hyperlink is grounded in a curated, version-controlled reference set — unsourced claims are dropped, not shown.
One topology, laptop to production
The same code runs on a laptop and in production through config swaps — no separate code paths to drift.
Store, storage, and executor modes flip between in-memory plus local files and Postgres plus blob storage plus a sandboxed runner; a single HTTP front door serves async jobs with live progress.
A joint effort
Built with
- Python
- Node.js
- TypeScript
- Next.js
- Anthropic
- Zod
- Playwright
- PostgreSQL
Math first, clinically conservative, reproducible by design
The edge isn’t content — it’s the forecast model and a deterministic allocation engine you can audit. The hybrid build keeps every plan trustworthy and resilient, even when the AI part fails.