Four attributes, one pass
Upload an image or a batch of images (extracted video frames work well). Neural Sentinel returns plate number, state, make/model/year, and color — each with a confidence score.
Tuned for tolling-style capture — angled plates, reflections, partial occlusion, mixed U.S. formats.
State inferred from plate artwork and layout. Covers U.S. state and DC plate designs, including many specialty issues.
Trim-level identification when possible. Reviewers can override the model if needed.
Normalized color field for filtering and search — not affected by plate enhancement filters.
A review queue, not a dumping ground
High-confidence detections auto-approve. Low-confidence detections land in a queue your team can burn through with the keyboard.
Confidence-gated auto-approval
Each vehicle is scored on plate, state, and vehicle attributes. Anything above your threshold is auto-approved and exported. Anything below routes to the queue — the share depends on your imagery and the threshold you set.
- Per-reviewer stats: total, auto-accepted, edited, rejected, escalated
- Average review time tracked per operator
- Supervisor view for claim history and escalations
- Approval rate and queue depth on every refresh
Enhance stubborn plates in seconds
When the plate is dark, glared, or low-contrast, your reviewer doesn't need Photoshop — the enhancements panel is built in.
Presets first, sliders when you need them
One-click presets — Auto, Hi-Contrast, Night, Deblur, Glare, Grayscale — cover the common cases. Fine-tune brightness, contrast, sharpness, saturation, exposure or gamma when a plate needs it. Enhancements apply only to the plate crop, so the vehicle shot stays true.
- Keyboard shortcuts for navigate, approve, reject, escalate
- Save & Approve commits extracted attributes in one action
- Manual override for any field with a non-obvious read
- Zoomable plate and vehicle crops
Every detail, one keystroke away
Every reviewed vehicle keeps the full record: the vehicle crop, the plate crop, extracted plate number and state with confidence, make/model/year, color, and detection metadata — frame index, timestamp, bounding box. Previous / Next moves you through the batch without a single mouse click.
- Edit any field; the audit trail preserves the original model output
- Detection metadata for timeline reconstruction and dispute handling
- Per-batch export: JSON, CSV, PDF
- Per-batch exports feed tolling, CAD/RMS, and evidence pipelines; REST integration available under contract
Under the hood
The pipeline, the people tiers, and what you get as a buyer.
Detection pipeline
- Proprietary detector for vehicle detection and cropping
- Fine-tuned vision-language model for plate, state, make/model, color
- State-inference fallback when the crop is low-quality
- Single-image or batch input · live WebSocket progress
- One VLM pass for all attributes — no separate OCR + classifier chain to maintain
People tiers
- Reviewer — confidence-gated queue, keyboard-first modal
- Supervisor — claim history, operator identity, time-per-item
- Org admin — users, thresholds, exports
- Platform admin — tenanting, roles, audit
- Every edit audited against the original model output
Training & data
- Fine-tuned on domain plate imagery for tolling-style capture conditions
- Angled plates, reflections, partial occlusion covered
- U.S. state and DC plate designs, including specialty and vanity formats
- PostgreSQL-backed with versioned Alembic migrations
- Your review edits can feed the next training cycle
Bring your review workload down
Pilot Neural Sentinel on a day's worth of your tolling imagery and measure how much of it auto-approves. Our team handles the integration.