Employee Recall — Persona Continuity with LoRA + RAG

An AI system that captures a departing employee’s writing style and historical knowledge using LoRA fine-tuning plus retrieval, enabling persona continuity for their successor.

Summary

When senior employees leave, their writing style and historical knowledge walk out the door with them. Employee Recall is an open-source methodology — and reference implementation — for capturing both as a small, locally-runnable AI model their successor can query.

The core idea: LoRA bakes the writing style into the model weights, RAG holds the memory in a vector index, and a system prompt anchors the identity. Trains in ~30 minutes for ~$0.25 on Colab, runs free on a Mac via Ollama.

Highlights

• Stack: Qwen2.5-7B base · Unsloth/PEFT/TRL for LoRA · BGE embeddings · FAISS · Ollama for local inference · FastAPI + n8n for Slack/REST integration.
• Reproducible synthetic corpus — 18,978 documents across two demo personas (Priya the CSM, Rohan the staff engineer), generated deterministically so no real PII is involved.
• Citations by default — answers reference [Source N] chunks from the retrieval index, so every claim has a trail back to a corpus filename.
• Spectrum of deployments — same pipeline supports anything from a low-risk “company RAG” through to a “departing employee twin” that needs serious privacy governance.

Links

• 💻 Repo: github.com/kader-xai/EmployeeRecall
• 📝 Blog write-up