01 / OPERATING CONTEXT
Tracking obligations across client correspondence.
A consulting agency may be waiting for information from one client, reviewing a decision with another and preparing a follow-up for a third. The next piece of evidence can arrive in a different email thread from the one where the obligation began.
Client Engagement keeps a durable record of the business issue across that changing correspondence. It connects new information with the existing client context, updates the working state and identifies the next action.
The operator can confirm an issue, adjust ownership, correct a proposed action or resolve the work. Those decisions remain part of the record when the agent processes subsequent evidence.
02 / IMPLEMENTATION SCOPE
Application and service capabilities.
Context across clients
Client workspaces and mailbox routing keep evidence connected to the appropriate engagement, supporting an agency working across concurrent client relationships.
Issues that outlast a thread
New communications can create an issue, add evidence to existing work, change responsibility or indicate that an obligation has been resolved.
Owned next actions
Separate suggested actions from confirmed work. Identify what needs follow-up, what is waiting on a client or third party, and what the operator needs to do next.
Operator control and continuity
Confirm, reject or edit proposed actions; assign an owner; resolve or suppress an issue. Explicit corrections survive subsequent agent runs.
03 / ARCHITECTURE & TOOLS
System boundaries and the stack behind them.
Microsoft 365 evidence and Google/Gemini reasoning connect through a custom workflow application. Persistent state and operator controls give the agent continuity across clients and recurring runs.
Evidence and client routing
Microsoft Graph ingestion and a connection registry associate approved evidence with a workspace and mailbox. Deterministic normalization prepares bounded evidence packets.
Reasoning and durable state
Gemini adjudicates the relationship between new evidence and existing issues. Validated updates flow into a durable read model containing issue state, evidence references and action history.
Operator interface and recurrence
An application interface and Admin API expose issues, suggested actions and controlled corrections. Scheduled processing maintains the working picture as new evidence arrives.
Tools, frameworks and their roles
- Microsoft Graph · Microsoft 365
- Read-only email evidence ingestion and client/workspace mailbox routing.
- Gemini · Google Gen AI SDK · Pydantic
- Issue adjudication, model integration and structured validation of evidence and decision contracts.
- Python · FastAPI · Uvicorn
- The processing pipeline and Admin API exposing controlled operator actions.
- React · TypeScript · Vite
- The operator interface for reviewing issues, confirming actions and correcting ownership or state.
- Firestore · Cloud Storage · Secret Manager
- Durable projections, stored run artifacts and runtime credential management.
- Cloud Run · Cloud Scheduler · Google IAP
- Application/runtime services, recurring processing and protected operator access.
- pytest · Ruff · Node.js test runner
- Pipeline and service tests, Python quality checks and frontend regression checks.
Domain logic, controls & engineering decisions
- The issue is a durable business record, with ownership and next action separate from the source email thread.
- The evidence pipeline separates normalization, candidate retrieval, bounded packet construction, model adjudication and validated state updates.
- Structured adjudication distinguishes a new issue, an update, a likely closure, reference-only information and work requiring human review.
- Deterministic candidate generation and fallback behavior support the reasoning path when model adjudication is unavailable.
- Explicit operator decisions remain authoritative across subsequent agent runs.
- Evidence references and action history preserve the basis for a proposed change or follow-up.
04 / WORKFLOW WALKTHROUGH
From new correspondence to a clear next action.
The business issue provides continuity while the evidence and the responsible person can change.
- 01
Collect the relevant evidence
Microsoft Graph brings approved email evidence into the appropriate client workspace. Normalization prepares it for issue matching and review.
OUTPUTClient-specific evidence - 02
Reason about the issue
Gemini considers the new evidence with the issue history to determine whether the work is new, has changed, needs action or may be resolved.
OUTPUTAn issue update and proposed next action - 03
Let the operator decide
The operator confirms or corrects the proposed action, assigns responsibility and records the disposition in the working interface.
OUTPUTConfirmed work with an owner - 04
Maintain the thread of responsibility
Later evidence updates the issue while retaining operator decisions and action history. Waiting work and unresolved obligations remain available for follow-through.
OUTPUTA continuing record across conversations
05 / SYSTEM OUTPUTS
What the system produces.
The agency gains a continuing view of client obligations, ownership and next actions. People can pick up the work with its history intact and move the engagement forward as new information arrives.
- Persistent client issues and evidence history
- Confirmed next actions and ownership
- Visible client and third-party dependencies
- Recorded resolutions and operator decisions