A personal ML interview preparation tool powered by Claude AI and RAG over your own ML textbooks.
| Book Exercises | Progress | Study Journal |
|---|---|---|
 |
 |
 |
- Claude-generated interview questions — every question is generated by Claude API for your topic, grounded in your books via RAG. No hardcoded question banks.
- Socratic evaluation — submit your answer, Claude evaluates it: what you got right, what's missing, a book reference, and an improved answer.
- Weak spot tracking — scores per topic tracked automatically. Random drill pulls 70% from your weakest topics.
- Book Exercises — real end-of-chapter exercises extracted from your PDFs with Claude model answers backed by book passages.
- Equation Cheat Sheet — key ML equations with Claude explanations grounded in your books.
- Study Journal — every answer and evaluation saved and reviewable.
- Framework Studio — learn OpenAI Agents SDK, CrewAI, and LangGraph by reading real project code with Claude explanations.
- Study Guide — structured 10-phase ML interview learning path with interview tips and direct drill links.
- Backend: Python + FastAPI + ChromaDB + PyMuPDF
- Frontend: React (single HTML file, no build step) + Tailwind CDN
- LLM: Claude API (
claude-sonnet-4-20250514) - Vector DB: ChromaDB (local, embedded)
- Package manager: UV
No LangChain. No LlamaIndex. Raw API calls only.
curl -LsSf https://astral.sh/uv/install.sh | shgit clone https://github.com/Mituvinci/ml-interview-drill.git
cd ml-interview-drill
uv synccp .env.example .env
# Edit .env — add your ANTHROPIC_API_KEYPlace PDFs in data/books/:
data/books/
├── goodfellow.pdf # Goodfellow — Deep Learning
├── chip_huyen.pdf # Chip Huyen — Designing ML Systems
└── geron.pdf # Géron — Hands-On ML
Any PDF works. Add more by editing BOOK_FILES in ingest.py.
You can also add markdown content (notes, Q&A books) via MARKDOWN_DIRS in ingest.py.
uv run python ingest.pyCreates a local ChromaDB vector store in data/chroma_db/.
uv run python extract_exercises.pyExtracts end-of-chapter exercises from your PDFs into data/book_exercises.json.
uv run uvicorn main:app --reload --port 8000ml-drill/
├── main.py # FastAPI app
├── ingest.py # PDF + Markdown → ChromaDB
├── extract_exercises.py # Extract book exercises from PDFs
├── api/ # All API endpoints
├── core/
│ ├── question_generator.py # Claude question generation + caching
│ ├── evaluator.py # Claude evaluation prompts
│ ├── rag.py # ChromaDB retrieval
│ └── ...
├── frontend/
│ ├── index.html # React app shell
│ └── components/ # React component files
├── data/
│ ├── books/ # Your PDFs go here (gitignored)
│ ├── study_guide.json # Learning path config
│ └── code_projects/ # Framework Studio code examples
└── notes/ # Personal study notes HTML (gitignored)
Every drill question is generated by Claude:
- You select a topic (e.g. "Adam Optimizer")
- App checks cache (
data/generated_questions.json) for unanswered questions on that topic - If none → calls Claude with a RAG-grounded prompt using your indexed books
- Question is cached — generated once, free forever after
The prompt produces Zoom-interview style questions: conceptual, comparison, explanation. Not "derive from scratch."
Place HTML files in notes/ and add the filename to ALLOWED_NOTES in main.py. Served at /notes/<filename>.
Set these secrets in your Space settings:
ANTHROPIC_API_KEY=your_key
HF_TOKEN=your_hf_token # optional — for syncing progress to HF Dataset
HF_DATASET_REPO=user/repo-name # optional
See core/hf_sync.py for the sync logic.
This tool is built to study from these excellent resources. No book content is distributed in this repository. You must obtain your own legal copies.
- Ian Goodfellow, Yoshua Bengio, Aaron Courville — Deep Learning (free online)
- Aurélien Géron — Hands-On Machine Learning with Scikit-Learn, Keras & TensorFlow (O'Reilly)
- Chip Huyen — Designing Machine Learning Systems (O'Reilly) and ML Interviews Book (free online)
- Ed Donner — LLM Engineering: Master AI & Large Language Models (Udemy) — the agentic AI module that inspired the Framework Studio feature
Place your own PDF copies in data/books/ before running ingest.py.
MIT