Local-first, privacy-centric personal assistant for macOS. Integrates Slack, Jira, Gmail, and arbitrary data sources via a generic connector framework. Automates task extraction and draft proposal through a Ghost Drafting workflow. All inferences default to local execution; cloud escalation is restricted and PII-redacted.
| Document | Description |
|---|---|
| UCE-Technical-Specifications-v1.1.md | Core spec: unified ingestion, air-gapping, relationship awareness, task extraction, Ghost Drafting |
| UCE-Technical-Specifications-v1.2-addendum.md | v1.2 additions: generic connector framework, PII-aware cloud LLM routing, on-demand UI generation, cloud search integration |
| UCE-Gap-Analysis.md | Gap analysis against existing Personal Assistant codebase — what exists, what's partial, what's absent, prioritized refactoring plan |
| EVALUATION-FINDINGS.md | Evaluation of v1.0 spec: 15 issues identified and resolved in v1.1 |
| CHANGES-SUMMARY.md | Diff summary: v1.0 → v1.1 |
- Local-first: All inference defaults to on-device (Apple Foundation Models + Phi-4/Llama via MLX). Cloud escalation requires explicit user confirmation.
- Privacy by partition: Data tagged
[Work],[Personal], or[Shared]at ingestion. Separate LanceDB indexes per partition.[Personal]never leaves the device. - PII redaction before cloud: Any text sent to a cloud LLM is redacted via
PIIRedactor(names, emails, phones, addresses replaced with consistent tokens; de-anonymized in response). - Connector-agnostic ingestion: Any data source implements a single
Connectorprotocol. Cross-cutting concerns (retry, dedup, auth, tagging) handled byConnectorHost— connector authors write source-specific logic only. - Human in the loop: Ghost Drafts require explicit approve/edit/reject before posting to any source system.