Municipal vehicles already drive every route in the network — refuse trucks, water tankers, traffic and law enforcement patrols. SignWatch turns their dashcams into a rolling sign inspection fleet: Gemini reads every frame for damaged, missing or obscured signage, Azure ML ranks what's actually dangerous, and a work order reaches a technician before a resident has to report a missing stop sign.
Road signage is one of the cheapest safety interventions a municipality maintains, and one of the least monitored. Most municipalities have no systematic inventory of sign condition — only a complaints line that catches a fraction of what's actually wrong.
A missing sign on a quiet residential street or a faded warning sign on a rural road may go unreported for months, even where the safety consequence is serious.
Walking or driving every route with a clipboard to inventory sign condition is expensive and infrequent — most municipalities can afford it once every few years at best.
A faded parking sign and a missing stop sign at an uncontrolled intersection are not the same risk, but without a prioritisation layer they compete equally for the same maintenance budget.
The operational loop behind every vehicle in the pilot fleet.
No single model does everything. Each provider sits behind a common adapter layer that records provider, model, task, timestamp and confidence with every result.
| Provider / layer | Job in SignWatch |
|---|---|
| AWS IoT Core | Handles telemetry from the in-vehicle gateway — GPS track, trip metadata and upload status for every route pass across the fleet. Not an AI model — the connectivity backbone. |
| Google Gemini (Vertex AI, latest) | Reads captured road frames to detect signs in view and classify each as intact, damaged, missing, faded, or obscured by vegetation or graffiti — the core computer vision layer. |
| Azure Machine Learning | Prioritises detections by safety criticality, weighting factors like sign type (stop, yield, pedestrian), intersection risk history and traffic volume where available. |
| GPT (OpenAI, current model) | Drafts the work order in plain language for the signage crew — what's wrong, where, and how urgently it needs attention. |
| Claude (Anthropic, latest production model) | Reviews recurring maintenance patterns across a location's history — a stop sign vandalised three times in a year is a different problem than a one-off, and Claude reads the maintenance log to surface that. |
| GPT Image | Generates correct-signage reference illustrations for the crew during reinstallation, always tagged AI-generated with the prompt summary and generation timestamp shown. |
Human-in-the-loop controls — Approve, Modify, Reject, Request inspection, Escalate — sit on every AI recommendation before it becomes a dispatched work order. Model identifiers are verified against current provider documentation at implementation time rather than hardcoded from memory.
Drag to rotate, scroll to zoom. Switch views to see each component on its own, or the full rig working together against the sign it's inspecting.
Problem. Road safety guidance in South Africa consistently identifies signage condition — visibility, reflectivity and presence at controlled intersections — as a maintainable risk factor, distinct from driver behaviour or road design. Municipal signage inventories are frequently out of date because inspection relies on ad hoc reporting rather than scheduled monitoring.
How the AI + IoT solution addresses it. SignWatch mounts a forward-facing camera and GPS unit on vehicles municipalities already operate on daily routes, converting normal municipal driving into a continuous, low-cost signage audit rather than a periodic manual one.
How the AI models integrate. Gemini classifies each sign detected in a route pass as intact, damaged, missing, faded, or obscured. Azure ML ranks detections so that a missing stop sign at an uncontrolled intersection is queued ahead of a faded parking restriction sign. GPT turns the ranked detection into a work order a signage crew can act on immediately, and Claude checks the location's maintenance history for recurring patterns worth flagging to a supervisor.
Evidence. The prioritisation logic in the demo model reflects publicly documented road-sign risk categories (regulatory and warning signs at intersections ranked above informational signage) rather than a live incident dataset — that distinction is labelled throughout the platform.
Problem. Sign retroreflectivity — how well a sign reflects headlights back to a driver — degrades gradually and is very difficult to assess from a single daytime glance, which is the only inspection most signs ever get. Traffic control device guidance treats reflectivity decline as a distinct maintenance category from physical damage.
How the AI + IoT solution addresses it. SignWatch's camera captures signs under normal driving conditions across different times of day as vehicles complete their routes, giving the detection model more than a single static daylight view to assess condition against.
How the AI models integrate. Gemini's classification includes a faded/low-reflectivity category distinct from physical damage, so a sign that is structurally fine but hard to see at night is queued differently than one that is bent or missing. Azure ML factors road type and speed limit into how urgently a reflectivity issue is treated, since a low-visibility warning sign matters more on a high-speed rural route than a slow residential street.
Evidence. The reflectivity-decline category in the demo model reflects general guidance from traffic control device literature on sign maintenance categories, not a calibrated photometric measurement from an installed sensor.
Filled cells are the active pilot fleet. Outlined cells are planned capacity, not signed municipal contracts.
SignWatch is built and operated by Tricloud Corp, a South African private company registered with the Companies and Intellectual Property Commission (CIPC).
Why we startedA damaged road sign rarely makes the news, until the day someone misses it. I started Tricloud because the people responsible for our roads should not have to wait for a crash or a complaint to learn a sign has faded, fallen or been hidden. The impact I want to make is simple: every municipality, big or small, knowing the condition of every sign it owns, and fixing the dangerous ones first.
Thuso Rankgoma Dammie
Director
Email: Director@tricloud.co.za
Phone: 061 885 2756
LinkedIn profile ↗
Tricloud Corp
Registration No: 2026 / 685596 / 07
Enterprise type: Private Company
Registered: 3 September 2026
Tax No: 9815880191
17 Saul Jacobs Street
Mindalore, Krugersdorp
Gauteng, 1739
South Africa
| Status | Municipality | Location | Condition | Confidence | Sign type | Last pass |
|---|
All 20 signs tracked across the pilot fleet. Real municipal and route names anchor the simulation to plausible South African locations; no live camera feed feeds these values yet.
| Status | Municipality | Location | Sign type | Condition | Confidence | Last pass |
|---|
Every AI finding shows the provider, model role, timestamp and confidence, and stays in a pending state until an operator approves, modifies, rejects, or escalates it. AI never certifies safety or triggers dispatch on its own.
Exportable in the production build as PDF/CSV. In this demo, the summary below reflects the current simulated session state.
| Signs currently flagged (any severity) | — |
| Safety-critical defects (session) | — |
| Work orders completed (session) | 0 |
| AI findings approved (session) | 0 |