NTO Operations Copilot
An evidence-grounded research coach for new Notice to Owner researchers.
Guide new Notice to Owner researchers through the complete operational process, retrieve work-order evidence, explain exceptions, and recommend the next approved action without replacing human verification.
Illustrative coaching exchange based on the documented GC-conflict procedure.
These counts describe the implemented synthetic dataset and interface. They are not accuracy, time-saving, or real-usage figures.
Project film · 2 min
A coach, not an autopilot.
The six research stages, the request queue, the read-only tool boundary, two-pass answers, the GC-conflict path, notice selection, and the cross-cloud architecture. Music only, with no narration; all work orders are synthetic.
Project overview
New NTO researchers have to work through property, recorded NOC, participant, and notice checks while customer intake, official records, and procedure don't always agree. The hard part is knowing which source to trust, what to record, and when to stop and ask.
NTO Operations Copilot guides the researcher through six stages, retrieves synthetic work-order evidence, explains exceptions, preserves uncertainty, and recommends the next approved action.
Key coaching capabilities
Process coaching
Explains the current research stage and the reason behind the next approved action.
Work-order grounding
Retrieves synthetic intake evidence through one allowlisted, read-only MCP tool.
Conflict preservation
Keeps customer claims separate from recorded evidence instead of silently choosing a winner.
Notice selection
References a state and project-class notice resource while requiring the underlying WO facts to be verified.
Human control
Stops on missing or conflicting evidence; the researcher stays responsible for verification and escalation.
Query-specific answers
Returns concise Answer, Why, and Next step guidance instead of repeating the full record.
System architecture
A cross-cloud evidence path with explicit policy boundaries.
The browser never talks to Foundry or holds credentials. The FastAPI gateway validates each request, calls the Foundry WO intake agent, and permits only the read-only get_work_order MCP tool, which reads synthetic records from AWS.
Evidence example
When the customer and the record disagree, both stay on the page.
A documented general-contractor conflict from the synthetic teaching set. The coach labels where each value came from and doesn't pick a winner.
Customer claim
Horizon Builders
General contractor as provided on intake. Unverified.
Recorded NOC evidence
Summit Construction
General contractor named on the recorded Notice of Commencement.
Approved conflict path
The approved procedure stays visible.
The coach points to the documented path rather than improvising one.
- 01
Preserve conflict
Record both names and their sources. Neither value overwrites the other.
- 02
Contact CC first
Follow the approved CC-first confirmation path.
- 03
Three calls, three emails
Document each attempt when the procedure requires it.
- 04
Return to customer
Send the issue back if the conflict remains unresolved.
- 05
Escalate when applicable
Route to human review when the operating procedure requires it.
Request queue
100 new research requests (SYN-WO-000121 to SYN-WO-000220) are grouped into 20 intake scenario families with five variants each. They arrive as ready for research, so no queued WO looks already reviewed, and a selection isn't marked loaded until the intake agent confirms the record.
Missing references show as Not provided. That is deliberate scenario data for the researcher to work through.
Notice selection
The coach's knowledge includes a state notice-selection reference. Work orders fall into four project/notice classes, and the researcher still verifies the state and class from the evidence before a notice is chosen.
An internal operational reference, not legal advice. Current rules and deadlines need verification.
Implementation
What is deployed
- WO validation and intake summary
- 100-item guided request queue
- Foundry and MCP evidence retrieval
- Answer, Why, and Next step responses
- Human review and escalation guidance
Tech stack
Built with
Assumptions & limitations
- The application uses synthetic work-order data only. It isn't connected to real customer workloads.
- The state notice-selection resource is an internal operational reference, not legal advice. Current rules and deadlines need verification.
- The coach recommends actions. It doesn't make autonomous legal or operational decisions, and it doesn't approve or release notices.
- No production accuracy, researcher-time reduction, or error-reduction figure has been measured. Establishing those would need a user study.