01 Capabilities
What we are actually engaged to build.
Three disciplines, each taken to the depth where we can be accountable for it in production. Most engagements draw on all three at once, because industrial problems rarely respect the boundary.
01 Discipline
Applied AI, ML & agentic systems
Models and agents that survive contact with the plant floor
- Harness first
- The evaluation harness is built before the model
- Agents constrained
- Sandboxed, authorised, traced and replayable
- Cost modelled
- Per-inference economics tracked as an SLO
Machine learning and agentic systems built as production software, not demonstrations — with data contracts, tool contracts, evaluation harnesses, drift detection and a cost envelope you can defend to a CFO.
Most AI initiatives fail somewhere between the promising prototype and the operational system. We work the other way around: define the decision the model or agent must improve, instrument the ground truth, then engineer the smallest system that moves that decision reliably. For agents specifically, the harness is the engineering — tool contracts, sandboxed execution, authorisation boundaries, trace capture and replay — because an agent that cannot be reviewed, reproduced or constrained has no place in an industrial environment. Everything ships with an evaluation suite, a rollback path and a unit economics model.
Representative work
- 01 Computer vision for inspection, metrology and defect classification
- 02 Predictive maintenance and anomaly detection on sensor and telemetry streams
- 03 Retrieval-augmented and agentic systems over technical corpora, SOPs and drawings
- 04 Agent harnesses: tool contracts, sandboxed execution, authorisation and audit trails
- 05 Evaluation harnesses and quality gates for models and agents alike
- 06 Forecasting and optimisation for throughput, yield, energy and scheduling
- 07 Edge inference on constrained industrial hardware, with fleet management
- 08 MLOps: feature stores, model registries, drift monitors, promotion gates
02 Discipline
AI-native transformation
Changing how the work works, not just the tooling
- Measured
- Baseline captured before any change lands
- Non-freezing
- Roadmap keeps moving during the migration
- Transferable
- Your team owns it when we step back
Operating-model change with an engineering spine — value-stream diagnostics, AI and agent-native ways of working, a defensible quality bar, and squads that measurably ship faster at the end of it.
Transformation programmes stall when they are run as change management with a software annex. We invert that. We start from the value stream and the constraint, define a target architecture that a real team can reach in quarters rather than years, and then build the paved road — templates, pipelines, agent tooling, quality gates, observability — that makes the new way of working the path of least resistance. Making delivery AI and agent-native is part of that road, not a parallel initiative: the operating model, the guardrails and the quality bar move together, and the effect is measured against a baseline captured beforehand.
Representative work
- 01 AI and agent-native ways of working across the delivery lifecycle
- 02 Agent pods: team topology, capability model, tooling stack and guardrails
- 03 Quality bar definition — gates, review standards and the measurement behind them
- 04 Value-stream and constraint diagnostics across OT, IT and engineering
- 05 Target architecture and a sequenced multi-quarter roadmap
- 06 Legacy decomposition, integration strategy and phased migration
- 07 Platform engineering: paved roads, golden paths, internal developer portals
- 08 Capability uplift — enablement, calibration and technical mentoring
03 Discipline
Product & platform engineering
Systems engineered for the decade, not the demo
- Observable
- SLOs, traces and dashboards ship with feature one
- Reversible
- Every release has a tested path backwards
- Documented
- Architecture decision records, not tribal memory
Cloud-native, event-driven, multi-region platforms and the products on top of them — designed around failure modes, observability and the total cost of running them.
We build the way regulated, latency-sensitive industries require: explicit failure modes, distributed consistency reasoned about rather than assumed, progressive delivery, and observability designed in from the first commit. The same rigour applies to the product surface — interfaces that engineers and operators actually want to use during a night shift.
Representative work
- 01 Greenfield product engineering — web, mobile, embedded and industrial UIs
- 02 Event-driven and streaming architectures for high-volume telemetry
- 03 Data platforms: lakehouse, time-series, semantic layers and governance
- 04 Kubernetes, service mesh, IaC, progressive delivery and DR strategy
- 05 Quality engineering as a product: automation, contract tests, load and chaos
- 06 Security engineering: threat modelling, supply chain, secrets, audit trails
04 Agentic systems
With agents, the harness is the engineering.
A capable model is now the commodity part. What decides whether an agent is allowed anywhere near an industrial or regulated process is the apparatus wrapped around it — and the bar it has to clear before it is promoted.
A The harness
-
Tool contracts
Every capability an agent can invoke is a typed, versioned, permissioned interface — not a free-form shell. If it is not in the contract, it cannot be reached.
-
Sandboxed execution
No ambient credentials, no unbounded network, no writes outside a declared scope. The blast radius of a bad decision is designed before the agent runs once.
-
Authorisation boundaries
Any action with a consequence passes an explicit gate — policy, approval or both. Read paths and write paths are never the same path.
-
Trace capture and replay
Every run is reconstructable: inputs, retrievals, tool calls, decisions and outputs. A disputed result can be replayed rather than argued about.
-
Budgets enforced, not advised
Token, cost, latency and action ceilings are enforced in the harness. Runaway loops are a design consideration, not an incident waiting to happen.
-
Contained failure
Timeouts, circuit breakers, and a tested path back to the previous behaviour. The fallback is rehearsed before go-live, not discovered during one.
B The quality bar
-
Task suites with agreed bars
Pass/fail thresholds defined before anything is built, against tasks drawn from how the work is actually done — including the awkward cases.
-
Slice reporting
Aggregate scores hide precisely the failures that destroy trust. Performance is reported per product variant, per shift, per site, per document type.
-
Regression gates in CI
Promotion is refused automatically when a slice regresses beyond tolerance. Model, prompt and tool changes are all treated as releases.
-
Adversarial suites
Prompt injection, tool misuse, data-exfiltration and over-permission paths tested as standard. Security review covers the agent, not only the API in front of it.
-
Sampled human adjudication
A known proportion of live decisions reviewed by a qualified human, producing fresh labelled data and an honest ongoing accuracy figure.
-
Drift on inputs and outcomes
Monitors on what goes in and on what the tools actually did — not just on the text that came out. Input drift precedes output degradation.
The same discipline applies whether the agent is assisting an operator, triaging maintenance history, or writing code inside your delivery pipeline. If it can take an action with a consequence, it passes a gate — and the gate is code, not a convention. How this runs in delivery.
05 Instruments
The toolchain, stated plainly.
We are not romantic about tools. These are the ones we know deeply enough to be accountable for in production.
Languages
- Go
- Python
- Java / Kotlin
- TypeScript
- Rust
- C++
AI & agents
- PyTorch
- ONNX / TensorRT
- scikit-learn
- Ray
- MCP tool servers
- Agent frameworks
- Eval harnesses
- Vector stores
Data
- Kafka / Pulsar
- Flink / Spark
- TimescaleDB
- Iceberg / Delta
- dbt
- DuckDB
Platform
- Kubernetes
- Terraform / Pulumi
- AWS · GCP · Azure
- ArgoCD
- Istio / Linkerd
- OpenTelemetry
Industrial
- OPC UA
- MQTT / Sparkplug B
- Modbus
- Edge gateways
- ROS 2
- PLC/SCADA integration
Scientific
- CUDA
- MPI / SLURM
- NumPy / SciPy stack
- Zarr / HDF5
- OpenCV
- Simulation toolchains
Working in a stack that is not listed here? Say so in the brief — we will tell you plainly whether we have the depth to be accountable for it in production, and where we would bring in a specialist rather than learn on your programme.
Next step
Bring the constraint. We will tell you what it really takes.
A 45-minute technical conversation. Bring the constraint you are stuck on and you will leave with an honest read on whether it is tractable, what it would take to move it, and how we would sequence the work.
or write directly — hello@bluestreaklabs.com