PRODUCT & ENGINEERING / OCTOBER 2026
From demonstration
to operational MVP.
A focused implementation plan for transmission-tower visual inspection and substation-transformer visual and thermal review.
1. Product boundary
The operational MVP supports evidence-based inspection review and maintenance prioritization. It does not control electrical infrastructure, issue autonomous switching commands, predict exact failure dates, or replace professional engineering review. Its first success criterion is a complete, traceable inspection journey.
2. Delivered prototype
The hosted MVP provides ChatGPT sign-in, organization workspaces, saved assets and inspection histories, original-image storage, human observation review, maintenance records, searchable text and PDF documents, and inspection reports. The AI image-analysis and evidence-linked Copilot integrations require a configured provider credential. They do not generate simulated results when that connection is absent. The recorded walkthrough shows the earlier illustrative prototype.
3. Core data model
| Record | Fields and relationships |
|---|---|
| Organization / Membership | Organization ID, identity subject, role, membership status. Administrator, inspector, reviewer, viewer. |
| Asset / Component | Organization ID, asset ID, type, location, manufacturer, commissioning date; components attached to assets. |
| Inspection / Media | Asset, capture date, inspector, type, processing state; original object key, checksum, capture metadata, derived preview. |
| Analysis / Observation | Inspection, component, evidence region, proposed type and priority, confidence semantics, model/prompt version. |
| Review / Score snapshot | Reviewer, decision, rationale, timestamp; rubric version and contributing confirmed observations. |
| Maintenance / Document | Asset, action/status, source reference, file version, extraction state, permission scope. |
| Report / Audit event | Immutable evidence references and decisions; actor, action, object, timestamp, correlation ID. |
4. Provider-independent architecture
Web application → API → Asset, Inspection, Observation, Review, Scoring, Reporting, and Retrieval services → PostgreSQL, object storage, job workers, and provider adapters.
Interfaces isolate identity, storage, queues, model inference, maps, and monitoring. AWS is one initial deployment candidate, rather than the product's core service. Container packaging and infrastructure definitions improve portability; alternative deployments still require adapter and security verification.
5. Inspection pipeline
- Authorize organization and asset; create a draft inspection.
- Issue scoped upload access. Validate format, size, checksum, and media safety; preserve originals.
- Queue analysis with an idempotency key. Expose uploaded, queued, processing, completed, and failed states.
- Extract capture metadata and generate previews. Route visual and supported radiometric inputs separately.
- Run evaluated inference and validate structured output. Retain evidence coordinates and model version.
- Present proposed observations; never silently confirm model output.
- Record reviewer decisions and rationale. Recompute versioned scores from confirmed findings.
- Export a review snapshot; retry failed jobs without duplicate observations.
6. Thermal and comparison constraints
Colorized thermal images allow visual-pattern review only. Measurements require supported radiometric data and calibration context. Comparisons must surface missing load, weather, distance, viewing angle, emissivity, and capture context. Image alignment exposes uncertainty; unmatched components must not become confident deterioration measurements.
7. AI evaluation before pilot use
Build a labeled dataset with engineering reviewers and documented data rights. Split evaluation by asset and campaign to limit leakage. Measure precision/recall by defect class, missed critical findings, false-alert burden, evidence localization, and reviewer agreement. Evaluate thermal outputs by format and capture context. Agree acceptance thresholds with pilot reviewers before use. No accuracy numbers are claimed today. Calibrate confidence before displaying percentages.
8. Copilot
Start with structured queries and permission-filtered retrieval. Apply organization scope before retrieval and when opening evidence. Responses include source IDs and document versions. Refuse unsupported conclusions; treat uploaded documents as untrusted data. Evaluate citation accuracy, retrieval quality, tenant isolation, and unsupported-answer rate. Copilot does not execute operational control actions.
9. Security and reliability
Operational requirements include authenticated memberships, least-privilege roles, tenant-scoped queries, scoped media access, encryption, secrets management, retention controls, backups and restore drills, review history, upload validation, job retries, monitoring, and incident response. Do not expose real utility coordinates in public demos. These are implementation requirements, not certifications claimed for this prototype.
10. Delivery stages
| Stage | Outcome | Acceptance gate |
|---|---|---|
| Presentation prototype | Website, connected sample workflow, recording, plan | Both journeys work; decisions update records; simulated outputs are labeled |
| Operational foundation | Identity, organizations, durable registry, uploads, jobs | Tenant isolation, persistence, failure recovery, restore verification |
| Visual pilot | Evaluated tower findings | Agreed accuracy thresholds, versioned outputs, reviewer traceability |
| Thermal + comparison | Supported formats, contextual comparisons | Measurement provenance, capture context, uncertainty |
| Retrieval + reports | Permission-aware Copilot, versioned reports, scoring | Citations, unsupported-answer handling, reproducible export |
| Utility pilot | Controlled engineering-partner deployment | Feedback, error analysis, security review, documented limits |
11. Pilot discovery and cost model
Confirm data ownership, imagery, asset identifiers, thermal formats, component taxonomy, review roles, existing processes, and integrations. Estimate original-media storage, previews, retention, inference per inspection, workers, retrieval, maps, and users. Instrument unit costs before commercial tiers. Schedule depends on dataset readiness and pilot scope.
12. Later roadmap
Vegetation intelligence, additional asset classes, work-order integrations, sensor context, and validated deterioration models follow the core inspection loop. Predictive maintenance requires longitudinal evidence and separate validation.
Return to Premvera ↗