Language models as drafting tools, with a teacher in the loop.

Synthetic case generation

Claude

Answer evaluation with rubric

Claude, plus retrieved context

PDF case detection and extraction

Claude, after PDFBox text extraction

Practice questions and evaluation

Claude

Image analysis and curation

Claude with image input

TR / EN translation

Claude

Gleason pattern suggestions

Claude (the grade is computed deterministically)

Embeddings

Ollama, nomic-embed-text

The Claude model is a configuration value. All calls are synchronous; there is no streaming.

Every Claude call passes through one client. It marks the system prompt for prompt caching, places image blocks before text in vision calls, logs input, output, cache-creation and cache-read tokens for each call, retries with exponential backoff where requested, and wraps all failures in one exception that the API returns as 502.

The evaluation prompt was restructured so the rubric and format rules form a static system prompt and only the case and answers vary. That ordering is what makes caching effective. No cost or latency numbers are claimed.

1. 01

   Chunk

   Case text is split into 500-character chunks with 100 characters of overlap.

2. 02

   Embed

   Each chunk is embedded by a local Ollama model and stored in MongoDB.

3. 03

   Rank

   The question title is embedded; cosine similarity is computed in memory over that case's chunks.

4. 04

   Prompt

   The top three chunks are added to the evaluation prompt. No chunks, no context.

Ranking in memory avoids running a vector database and is adequate for the few chunks a single case produces. It is not meant for corpus-wide search, and it was not benchmarked against alternatives.

1. 01

   Extract text

   PDFBox reads the PDF. Non-admin teachers are limited to 15,000 characters.

2. 02

   Detect cases

   A first Claude call returns how many cases there are and their character ranges.

3. 03

   Extract fields

   Each range gets its own call that returns structured case fields.

4. 04

   Store and index

   Cases are saved, images go to GridFS, and text is chunked and embedded.

Image analysis produces an educational description of a case image and is stored as pending until a teacher approves it. Images are downscaled, and can be cropped, before being sent. Tumor grades are computed by pure, unit-tested calculators; the model may only suggest patterns, and each grading records whether it was manual or AI-suggested.

Streaming responses, per-user rate limits on AI endpoints, schema validation beyond JSON parsing, and a measured cost or latency baseline. An evaluation harness skeleton with synthetic cases exists; no benchmark results are claimed.