DHLM

The context behind a decision is usually split across an inbox, a calendar, and a folder of files.

One email holds a detail, a calendar event holds the time, and a document holds the rest. Answering a simple question about the week means opening three apps and stitching the answer together by hand.

Handing a general AI assistant full access to all of it swaps a searching problem for a trust problem: no clear limit on what it reads, and no clear record of what it changed.

Build one private workspace that fetches only what a question needs, cites it, and changes nothing without confirmation.

A rule-based router reads each request and decides which sources it needs before any model is called. Gmail, Calendar, and Drive are fetched in parallel with read-only access, and personal notes come from a ranked search over a Markdown knowledge base. Gemini streams the answer with knowledge and live-data citations kept visually separate.

The site at dhlm.davishiggins.com is a public case study built on synthetic data. The application itself sits behind a Google sign-in that admits one authorized account.

1. 01Chat and command workspace with streamed answers and source citations beside each response
2. 02Deterministic request router that selects notes, Gmail, Calendar, or Drive for each question
3. 03Google integrations for Gmail search, upcoming Calendar events, and Drive Docs, Sheets, Slides, and text files
4. 04Two-step draft and event creation: prepare, preview, then confirm with a single-use token
5. 05Memory updates proposed from conversation and written only after approval
6. 06Career workspace for job discovery, application tracking, and labelling recruiter email
7. 07Public case study with synthetic previews of six workflows, the request architecture, and access controls

• Sign-in requests a basic profile only. Gmail, Calendar, and Drive are connected in a separate, read-only consent step.
• Gmail drafts and Calendar events are optional, separately granted, and off by default. DHLM has no send permission, so sending email stays manual.
• Each draft or event is bound to a five-minute, single-use confirmation. Change the details after review and the action is refused.
• Gemini runs on the server only, with a fast default model and a deeper one used on request within a daily budget.
• The Google refresh token is sealed with AES-256-GCM in an HTTP-only cookie rather than stored in the database.
• Sign-ins, connections, and every prepared or executed write are recorded in an audit log.
• The morning brief runs when asked. The only scheduled job handles job discovery and data retention, and never touches Gmail or Calendar.
• There is no public sign-up. The case study shares no interface, data, or links with the private application.