Multimodal operations · 2026
CrewVerify
Use AI for difficult visual transcription, deterministic code for observable checks and people for accountable approval.
I defined the product boundary and built the authenticated extraction route, validation engine, review stages and audit-oriented interface.
- Responsibility
- Product definition + full-stack build
- Core stack
- Next.js 16 · React 19 · Claude multimodal
- Current scope
- Functional assessment prototype
The operating problem
Paper timesheets remain practical on construction sites, but office review is repetitive and errors can affect payroll preparation, job costing and accountability.
Evidence
What exists in the current build.
These numbers describe tested product scope—not production impact.
Inside the build
Screenshots of the running product.
Captured from the application itself—no mockups, no rendered concepts.
System workflow
From input to accountable action.
The interface keeps model work, deterministic checks and human responsibility distinct.
Capture
Start from the photographed field record.
A reviewer selects a synthetic sample or uploads a timesheet image through an authenticated application route.
The demo samples exercise complete, ambiguous and duplicate-entry scenarios.
Extract
Ask the model to transcribe—not decide.
The vision model returns schema-constrained attendance fields and leaves unreadable values empty instead of guessing.
The model does not match employees, calculate hours or approve a record.
Validate
Route exceptions using observable facts.
Code checks required fields, unique roster matches, title alignment, hours differences, duplicate rows and long shifts.
Routing does not depend on a model’s self-reported confidence.
Approve
Keep consequential decisions with people.
A foreman resolves row-level exceptions and a manager provides a separate final sign-off before export.
Excluded records remain visible and every correction requires an explicit action.
01 / Product thesis
A narrow architecture is more credible than broad autonomy.
Probabilistic perception is used only where it is valuable. Inspectable rules handle routing, and named human owners retain responsibility for judgment.
This separation makes the workflow easier to test and prevents model confidence from being mistaken for evidence.
02 / Safety boundary
The product deliberately stops before payroll judgment.
CrewVerify does not certify payroll, interpret agreements or make legal determinations.
Uploaded images are processed server-side, credentials never reach the browser and AI output cannot approve its own work.
Design principles
- Transcribe only visible evidence.
- Never silently repair the source record.
- Require explicit duplicate decisions.
- Preserve separate foreman and manager responsibility.