Edge Runtime
Sub-frame inference on commodity NVRs.
- 12ms P99
- Multi-camera scheduler
- Offline-tolerant
v4.2 · Autonomous vision agents
The agentic layer
for the physical world.
Thinking Stack turns every camera into a reasoning system. We deploy AI agents that see what happens on your floor, understand context, and act — before incidents become losses.
- <1sec
- Edge inference
- 6
- Industries
- 24/7
- Autonomous ops
worker · ppe ok
forklift · zone b
Live inference
38.42° N / 27.14° E
Thinking StackSee · Understand · Act
Camera feed · Zone B
// Affiliations
- NVIDIA Inception
- Microsoft for Startups
- Google for Startups
- AWS Activate
- Azure Marketplace
- Intel Partner Alliance
- NASSCOM
- Raspberry Pi
// Industries we operate in
One Platform, Many physical worlds.
The same reasoning core, tuned for the constraints and vocabulary of each domain. Pick your floor.
01 / Workplace safety
Stop incidents before they happen.
Vision agents watch every floor for PPE compliance, restricted zones, and near-miss patterns — then trigger the right response in under a second.
- Real-time PPE and hazard detection across shifts
- Escalation to supervisors via SMS, Teams or Slack
- Incident reports generated automatically
Safety · agent active
02 / Physical security
A guard that never blinks.
Perimeter, access points and after-hours activity monitored continuously — with intrusion, loitering and tailgating escalated the moment they happen, not the morning after.
- Intrusion, loitering and perimeter-breach detection
- Tailgating and unauthorised access at controlled doors
- Verified alerts with clip evidence to guard rooms and VMS
Security · agent active
03 / Retail intelligence
Every store, staffed by a silent analyst.
Shelf compliance, footfall attribution, and shrink prevention — running on the cameras you already have.
- Planogram and out-of-stock detection
- Conversion by aisle, hour and staff
- Loss-prevention agents that flag anomalies
Retail · agent active
04 / Manufacturing
A quality inspector on every line.
Detect defects, monitor cycle time, and correlate downtime causes across lines — without ripping out your MES.
- Sub-second defect classification at the edge
- Cycle-time and OEE observability
- Root-cause traces linked to video evidence
Manufacturing · agent active
05 / Logistics
See the yard, the dock, the trailer.
Track loading operations, trailer occupancy, and forklift routes — closing the visibility gap between WMS and reality.
- Dock-door utilization and dwell
- Damage capture at load and unload
- Route safety and near-miss telemetry
Logistics · agent active
06 / Cities
Public spaces that respond in real time.
Deploy vision agents on existing municipal cameras to monitor traffic, waste, incidents and pedestrian safety — with human oversight built in.
- Traffic and incident classification
- Illegal dumping and infrastructure alerts
- Auditable decisions with operator sign-off
Cities · agent active
// How it works
Moving beyond detection to autonomous action
AI detection is only the first step. The real value is an autonomous layer that perceives, understands context, decides, and acts — closing the loop without a human in the middle.
scroll →
01
Perceive
See everything, everywhere
Detect entities, events and anomalies across any visual input — indoor, outdoor, thermal, PTZ — at 12ms per frame.
YOLO-derivedMulti-modalEdge-first
02
Reason
Understand the situation
Fuse multi-camera, multi-modal signals into real context — who, what, where, and why it matters.
Scene graphsDomain LLMContext fusion
03
Decide
Choose the response
Apply rules, risk and intent — explainable and auditable, with human-in-the-loop wherever you need it.
Policy engineRisk scoringHuman-in-loop
04
Act
Close the loop
Trigger workflows and systems — alert, ticket, halt, dispatch — then learn from the outcome.
WebhooksSIEM / MESFeedback loop
// Product
The stack, module by module.
Adopt the whole platform or bring your own components. Every module speaks open protocols.
Reasoning Core
A domain LLM that grounds vision in your ops.
- Scene graphs
- Site-specific SOPs
- Explainable outputs
Agent Fabric
Composable agents for every incident type.
- Notify · Ticket · Halt
- Human-in-the-loop
- Full audit trail
Policy Engine
Codify the rules your floor already runs on.
- Zone & shift logic
- PPE & access
- Regulatory presets
Ops Console
One pane of glass for every site.
- Live feeds
- Incident timeline
- Team assignment
Integrations
Native connectors, no glue code required.
- SIEM · MES · WMS
- Slack · Teams · PagerDuty
- REST · webhooks
// Showcase
Agents running on real footage.
Live detections from deployed sites — every box, count, and alert produced by the same stack you'd run on your cameras.
Loading SOP Monitoring
Manufacturing & WarehouseLive
Intrusion Detection
Safety & SurveillanceLive
Spill Detection
RetailLive
Emotion Analysis
RetailLive
Customer Journey Maps
RetailLive
Safety PPE Check
SafetyLive
// Journal
Research & field notes.
research·2026 · 08
Grounding vision agents in operational language
How we move from raw detections to agents that understand SOPs, alerts, and the everyday language of the operators who run the site.
Read the piece →
research·2026 · 08
12ms at the edge: designing for constrained hardware
Why the most useful vision agents run on the cameras, not in the cloud — and how we design models that stay under 12ms on a factory edge node.
Read the piece →
// About
A team building the reasoning layer for the physical world.
Thinking Stack is a small, senior team of AI researchers, distributed-systems engineers, and operations model designers. We shipped AI into factories, retail chains and cities long before it was a category.
We believe the next decade of AI won't live in a chat window, it will live inside the cameras, sensors and vehicles that make the physical economy run. That's the stack we're building.
2019
Founded