Technology
The stack, per practice, plus the cross-cutting plumbing.
Open-source first across every practice. We adopt commercial software only where the licence buys something we'd otherwise engineer ourselves. Each section below is what we actually run — not what we could run if a project asked for it.
Per-practice stack
What runs in each practice.
Four practice-specific stacks. The fifth practice — AI-native film production — composes the generative-media stack with editorial integrations; we list those under media.
Safety & CV
Read the practice →Edge inference
- NVIDIA Triton Inference Server
- ONNX Runtime
- TensorRT
- CUDA 12.x
Streaming
- MediaMTX (RTSP/RTMP/SRT router)
- GStreamer pipelines
- ONVIF profile-S/T cameras
Data plane
- Apache Kafka
- PostgreSQL 16
- MinIO (S3-compatible)
- Redis
Identity
- Keycloak with OIDC/SAML federation
- SCIM provisioning to client IdP
Application
- React + TypeScript Ops Console
- Node 20 LTS services
Agentic Systems
Read the practice →Orchestration
- LangGraph
- CrewAI
- Temporal (long-running workflows)
- Custom Python/TS runtimes
Models
- Anthropic Claude
- OpenAI GPT family
- Google Gemini
- Open-weights via vLLM / TGI / Ollama
Tool integration
- Model Context Protocol (MCP) servers + clients
- OpenAPI-derived tool wrappers
- Auth-aware connectors (OAuth, mTLS, signed-JWT)
Memory & retrieval
- Postgres + pgvector
- Qdrant
- Redis for short-term state
- OpenSearch for hybrid retrieval
Eval & observability
- Langfuse / Helicone / custom dashboards
- Promptfoo / DeepEval / bespoke eval harnesses
- Structured tracing via OpenTelemetry
Cost governance
- Per-team and per-workflow token budgets
- Circuit breakers on cost spikes
- Per-model fallback policies
Generative Media
Read the practice →Generative model runtimes
- ComfyUI workflow runtime
- Stable Diffusion XL, Flux, Wan 2.x, Stable Video Diffusion
- Custom LoRA / adapter training (per show, per asset)
DCC integrations
- Nuke gizmos (Python + Blink)
- Houdini HDAs
- Maya plugins
- Blender add-ons
Pipeline & orchestration
- Deadline / OpenCue / Tractor render farms
- Pixar USD for asset interchange
- OpenColorIO (ACES) for colour
- S3-compatible object stores (MinIO on-prem, AWS S3 in cloud-burst)
Editorial & post
- Avid Media Composer, DaVinci Resolve, Adobe Premiere integrations
- Per-language voice / dubbing toolchains
- Unreal Engine 5 for virtual production
Provenance
- Per-asset model + prompt + approval metadata
- Immutable consent and approvals log
- C2PA-compatible content credentials where required
Custom ML & Research
Read the practice →Training & adaptation
- PyTorch
- JAX where appropriate
- HuggingFace ecosystem
- Ray for distributed training
Experiment & data
- Weights & Biases
- DVC and Git LFS for data versioning
- Bespoke eval harnesses per project
Deployment
- ONNX, TensorRT, GGUF for edge artefacts
- vLLM / TGI for self-hosted inference
- Triton for general serving
Cross-cutting
What every practice depends on.
Observability, delivery, and security baselines that apply regardless of practice. We don't ship without these.
Observability
- Prometheus + Grafana
- OpenTelemetry traces
- Loki for log aggregation
Delivery
- Docker + docker-compose
- Ansible for fleet config
- Git-flow with environment promotion
- GitHub Actions / GitLab CI
Security baseline
- mTLS service-to-service
- Vault / SOPS for secrets
- Per-environment IAM scoped to least privilege
Reference architecture · Safety practice
Edge inference, integration plane, central NOC.
Three lanes, left-to-right data flow. This is the safety-vision reference — the practice where edge autonomy is non-negotiable. Agentic and media practices use different topologies; ask during scoping.
Hardware OEMs · Safety practice
Who we specify for safety-vision deployments, and how we work with you on procurement.
Default model: we specify; you procure directly from the OEM. No mark-up, no working-capital lock-in. We also handle hardware sales when a client prefers single-vendor procurement — commercial terms agreed transparently in Discovery. Either way, we pre-verify every part against the design spec when it lands. Other practices have minimal hardware footprint — agentic and media workloads run on the client's existing GPU compute or cloud, sized during Discovery.
Compute & GPU appliances
GPU class: NVIDIA RTX PRO 6000 / L40S / L4 — sized to camera count, module mix, and frame rate.
Cameras and NVRs
We are camera-agnostic — your existing estate is the starting point. Only when modules need higher pixel density or specific anti-spoofing do we recommend replacements.
FRS terminals
Depth + IR anti-spoofing is mandatory at perimeter terminals. We do not deploy 2D-only terminals at gates.
Networking and access
Camera segregation onto a dedicated VLAN with policy-based isolation from corporate LAN is baseline.
Audio analytics
All audio analytics inference is local. No cloud upload of audio.
Want the full reference architecture document?
The detailed reference architecture — including sizing tables, network design, and integration patterns — is shared during Discovery.