Night Shift – Local Agentic Recruiter Outreach

Status: Under active development. The initial skeleton already runs end-to-end on Apple Silicon; the public dashboard is being polished.

The Problem

Recruiting inboxes accumulate years of outreach that mixes genuine opportunities with automated blasts. Traditional tooling requires exporting data to SaaS platforms, breaking privacy promises and making iteration slow. I wanted a local workflow where inference, embeddings, and storage never leave my laptop yet still act autonomously.

System Architecture

Night Shift follows a Lambda-style design:

1. Batch Layer (Cold Start) – Imports .mbox exports (~27k emails) and passes them through a janitor pipeline that strips signatures, HTML cruft, and disclaimers before vectorization.
2. Speed Layer (Live Sync) – An Azure AD public client (scoped to Mail.Read) fetches incremental updates behind enterprise firewalls and keeps drafts in sync.
3. Agentic Brain (LangGraph) – Replaces brittle sequential chains with an asynchronous state machine:
   • Classifier Gatekeeper – Quantized Llama 3.1 JSON mode separates genuine intent from spam.
   • Historian – Local FAISS store recalls previous dialogs with the same firm or recruiter.
   • Drafter – Synthesizes a reply using the candidate resume, recruiter context, and retrieved history.
4. Human-in-the-Loop UI – Streamlit dashboard surfaces drafts for one-click approval or manual edits before anything is sent.

Local Stack

Layer	Tech	Why
Compute	Ollama (Llama 3.1, nomic-embed-text)	On-device inference eliminates API cost and keeps PII local.
Orchestration	LangGraph	Explicit state + interrupts prevent runaway chains and enable partial retries.
Storage	FAISS + SQLite	Fast enough for <100k vectors, no cloud dependencies.
Ingestion	Native Python parsers	Direct .mbox access bypasses throttled APIs and corporate controls.

Flow

1. python -m src.ingest_local cleans and stages historical mailboxes.
2. python -m src.vectorize builds embeddings into a git-ignored FAISS index.
3. streamlit run interface/dashboard.py launches the reviewer surface (currently in progress).

All components run on Apple Silicon with 4-bit quantization; end-to-end draft generation completes in under 15 seconds.

Security Posture

• Zero Data Exfiltration: No cloud inference, no third-party storage.
• Scoped OAuth: The Azure client only reads mail and drafts.
• Local Secrets: .env holds runtime credentials; data/ remains git-ignored.

Roadmap

• Surface per-company histories directly inside the Streamlit UI.
• Add “auto-send” with configurable guardrails.
• Package the ingest/vectorize steps as reusable CLI commands.

---

Repository: github.com/nnigam96/night-shift