Fine-Tuning · Retrieval
Resume Griller
Practice interviews with resume-specific questions, gap detection, and adaptive follow-ups using retrieval and a fine-tuned language model.
- Python
- FastAPI
- Mistral-7B
- LoRA/PEFT
- LangGraph
- ChromaDB
- vLLM
- Next.js
- GCP
What I built
Resume-aware interviews with 18 gap types and inspectable, persistent session state.
- Fine-tuned Mistral-7B with LoRA/PEFT for interview evaluation and follow-up behavior, grounded in a ChromaDB resume retrieval layer.
- Separated Groq preprocessing from custom-model interview execution, with vLLM on GCP for the LoRA inference path.
- Used a LangGraph interview flow with SQLite checkpoints, multiple interview modes, and optional voice support.
How it works
- Parse and semantically chunk a resume, then embed it in ChromaDB.
- Prepare a compact resume summary and interview questions.
- Evaluate each answer, identify gaps, and generate a contextual follow-up.
- Checkpoint session state and return the next turn over the API/WebSocket interface.
Evaluation & evidence
The repository documents 18 gap categories and seven scoring dimensions. A reproducible end-to-end latency or interviewer-quality benchmark is not published here.
Scope & limitations
The custom inference path needs a configured GPU service and model adapter. Available provider integrations do not establish that a public hosted demo is running.