Cross-practice principles.
These hold whether the practice is safety vision, agentic systems, generative media, film production, or custom ML. The buyer changes. The stack changes. The SLO changes. The engineering posture does not.
Engineered for production
Every project clears a defined measurement bar before deployment. Safety vision: per-module accuracy SLAs. Agentic systems: task-success rates and cost-per-task ceilings. VFX: supervisor accept-rates. Film production: signed approvals at every milestone. Demonstration quality is not a production threshold.
Evaluation before model
The evaluation framework is defined before the model is trained or selected. If success criteria cannot be characterised at the outset, they cannot be recognised at delivery. This discipline is the single highest-leverage choice across every practice.
Deployed on the right infrastructure
On-premise and edge deployments for safety vision and enterprise agentic work where data sensitivity and latency require it. Cloud or hybrid topology for VFX render bursts and film post where elastic scale is the appropriate fit. The architecture follows the workload, not the vendor's licensing model.
Defined responsibility boundaries
Each practice publishes the categories we will not build — mass surveillance, non-consensual likeness, undisclosed synthetic content, fully autonomous agents on critical operations. The list is contractual, not aspirational. We share it during Discovery.
Observable, reversible operations
Every consequential decision is logged with enough context to reconstruct it on review. Where architecture permits, the action is reversible — a human undoes what an agent did, a supervisor rejects a generated shot, an operator recalls an alert.
The sections below go deep on how these principles play out in our lead practice — safety & computer vision. The other practice pages each carry their own engineering notes: agentic systems, VFX pipeline, film production, custom ML.
Platform architecture, not per-site rebuilds.
The default in the safety-vision market is to bid every site fresh: re-configure the VMS, re-train the models, re-deploy the dashboards. It's profitable for the vendor and expensive for the buyer. We rejected that approach early.
Instead, we build the platform once — edge inference engine, Ops Console, identity broker, integration adapters, alert routing — and treat each new site as a configuration exercise on top of that platform. A site profile defines the camera grid, the active modules, the SLA bands, the integration endpoints, and the compliance posture. Going from one site to twenty is editing configuration, not writing software.
Concretely, this is where the 60–75% per-site cost reduction comes from. The platform investment is front-loaded into the Discovery and platform-build stages; each subsequent site amortises against it. You see the per-site cost curve flatten dramatically after site three.
Honest per-module accuracy SLAs.
Accuracy in computer vision is not one number. It depends on lighting, camera angle, target distance, motion blur, occlusion, and the specific class of event. A blanket "99% accuracy" claim is almost always a marketing artefact — it tells the buyer nothing useful.
We commit to per-module accuracy bands in the contract, measured on the client's actual cameras and actual events. The bands are honest engineering ceilings — what we have repeatedly delivered in production, not what we hope to deliver. Examples from our standard SLA template:
- Indoor adult face recognition under controlled lighting — 95%
- Under-12 face recognition with periodic re-enrolment — 90%
- Weapon detection on P1 events (visible firearms and large blades) — 90%
- Behavioural events (loitering, intrusion in defined polygons) — 85%
- ANPR on standard plates, daylight — 94%; aged or night — 85–88%
Edge-first inference, on your hardware.
We run inference on a GPU appliance at the client's site. Video does not leave the network for AI processing — alerts, metadata, and clip references can be replicated to a central NOC, but raw footage stays put unless the client explicitly contracts otherwise.
The benefits compound. Latency stays under 100ms because there is no round trip. WAN failures don't take the site offline — modules keep running, alerts keep firing, the operator console at the site stays live. There is no per-feature licensing meter ticking against you in someone's billing system. Your data sovereignty is intact under the DPDP Act because the data never crossed an authorisation boundary you didn't approve.
For multi-site clients, a small central NOC pulls events and dashboards over a controlled channel. The NOC is a window into the sites, not the brain of the sites.
DPDP Act 2023 compliance, by design.
Every engagement starts with a Data Protection Impact Assessment. The DPIA is not a checkbox; it is the document that decides what we will and will not deploy at your site. It covers:
- Lawful basis for each category of processing (employee, customer, child, patient)
- Verifiable parental consent flows for any under-18 data — particularly biometrics
- Data minimisation: we capture only what the deployed module needs to function
- Retention schedules — 30-day hot, then policy-driven cold storage or deletion
- Right-to-erasure workflows for individuals, with audit log of the erasure itself
- Cross-border transfer assessment — defaulting to "no transfer" unless explicitly approved
The DPIA framework is sharable. If you are running a procurement, we will provide our DPIA template for your legal and compliance teams to review before any commercial commitment.
Continuous ML lifecycle, not a frozen model.
A model deployed once and never touched degrades. Lighting changes seasonally, people age, camera mounts drift, new demographics enter the site. We commit to a quarterly retraining cadence on every module under maintenance, with model versioning that lets us roll back if a new version regresses on any subgroup.
For face recognition specifically, we run per-site accuracy tuning every quarter — the same global model can produce different per-site numbers because the camera estate, the demographics, and the lighting differ. Tuning closes that gap. For K-12 face recognition, an annual re-enrolment is part of the calendar, baked into the school's academic-year start.
Hardware advisory by default. Sales available when the client prefers it.
We specify the hardware: camera models, GPU appliances, FRS terminals, NVR/MediaMTX configurations, network gear, identity-broker boxes. The exact OEMs we work with are listed on the Technology page.
Default: advisory.
The client procures directly from approved OEMs or distributors. We do not take mark-up on hardware. We do not introduce working-capital lock-in. We do not carry FX risk between an order in dollars and a quote in rupees. This is the model we recommend — it produces the cleanest commercial terms and the most defensible bill of materials for the client's CFO.
Either way, we provide pre-purchase verification of every part against the design spec, factory-fresh configuration when the hardware lands, and a procurement support window where we sit alongside your team in vendor negotiations.
On request: sales.
When a client prefers single-vendor procurement — one purchase order, one accountable party for both the platform and the underlying hardware — we will handle the sales path as well. Commercial terms, including any administrative margin on hardware, are agreed transparently in Discovery. There are no hidden mark-ups regardless of which model the client chooses.