Evangeline is a robotics consultancy and bespoke manufacturing partner. We work with operators, integrators, and enterprises deploying humanoid and industrial robotics that must be secure, useful, and defensible when they meet the physical world.
We work across four tightly integrated practice areas. Most engagements combine two or more. All start with a scoping conversation and a short written proposal.
Autonomous physical systems inherit every cyber-physical risk of the operational technology world, plus new classes of risk introduced by learning models and remote update pipelines. We design and audit device identity, edge compute isolation, telemetry integrity, model containment, and incident response for organisations deploying at scale.
Programs that integrate learning models with real actuators. Scoped, built, and handed off.
Prototype to run-rate. Housing, subsystems, and full-stack builds through vetted partner shops.
Average engagement horizon from cybersecurity readiness assessment to production deployment.
Purpose-built data capture for humanoid and industrial robotics: task-specific demonstrations, sensor packs, environmental variation, edge cases. Delivered under NDA with defined ownership and licensing terms.
Every autonomous physical system carries an identity, an operating perimeter, a failure mode, and an accountability chain. Evangeline designs those before we design the code.
The narrower your task and the higher its consequence, the more asymmetric the value of a well-scoped custom data set becomes. We build data sets that a general model cannot replicate.
We work across sectors where physical AI has real operational stakes. These are the environments we know deeply. If you are working somewhere adjacent, ask us.
Fleet coordination, warehouse humanoid deployment, and dock automation at run-rate scale.
S/02Surgical assistance, hospital logistics robots, pharmacy automation, and lab robotics.
S/03Cobot cells, precision assembly, quality inspection, and human-plus-robot line design.
S/04Front-of-house service robots, back-of-house automation, and guest-facing deployment ethics.
S/05Non-weapon autonomous systems: logistics, search, environmental sensing, base infrastructure.
S/06In-store robotics, autonomous last-mile delivery pilots, and store-format transformation.
Physical AI systems fail in ways software-only systems do not. A misconfigured update, a spoofed sensor stream, or a compromised model can produce physical consequence at production speed. Evangeline treats cybersecurity as a design discipline, not an audit checkbox.
Our cybersecurity engagements typically cover:
Programs that put learning models on real hardware. Scoped to a defined task, delivered as production-ready software with the operational documentation and handoff to sustain it internally.
Typical engagements combine three layers: model, integration, and operator interface. We build across all three or slot into whichever your team has capacity gaps against.
Programs are scoped in weeks, delivered in months, and handed off with the documentation and tooling required for your team to own them.
Task-specific data sets for humanoid and industrial robotics. Purpose-built, licensed on terms you can defend, delivered with the provenance record you need for audit and regulatory review.
General-purpose models rarely produce the narrow, high-consequence competence a production deployment requires. Custom data sets close the gap.
Typical deliverables include:
Prototype through low-volume production, coordinated through a network of vetted manufacturing partners. Housings, subsystems, custom actuators, and end-of-arm tooling built to the specification, not to the catalogue.
Manufacturing engagements typically begin at prototype and follow the same hardware through iteration into low-volume production. Coordination, quality, and IP protection sit with Evangeline; execution sits with the shop best suited to the part.
The bottleneck for physical AI is not the model. It is the operational discipline around it.
Evangeline · Practice note, Q3 MMXXVIWorking notes from our manipulation data practice: what we have learned running capture programmes, written for the people who buy, build, and audit training data. No gating, no forms.
First-person capture, why head and wrist views carry different information, and how episodes end up in RLDS.
Read the note → FN/02Self-occlusion, infinite states, resets that never repeat. Why cloth and garment data is hard, and priced accordingly.
Read the note → FN/03Task cards, rigs, synchronisation, episode segmentation, QC, and labelling. Our pipeline, end to end.
Read the note → FN/04Two ways to produce manipulation data, failing in different ways. The honest tradeoffs, including against our own interest.
Read the note → FN/05Sync accuracy, label taxonomy, outcome grading, capture hygiene, and the acceptance test every buyer should run.
Read the note → FN/06Fastening, folding, pouring, tray service. The frontier tasks robots keep bouncing off, and what capture must include.
Read the note →These sit alongside our four disciplines. Often bundled into a larger engagement, sometimes scoped standalone when the fit is clear.
ISO 10218, ISO/TS 15066, CE marking pathways for autonomous and collaborative systems. Gap analysis and technical file support.
A/02Simulation pipeline design, domain-randomisation strategy, and deployment validation for models trained partially or entirely in simulation.
A/03Remote supervision architecture, fleet orchestration, incident escalation paths, and the operational unit economics that make it sustainable.
A/04LiDAR, structured light, stereo vision, IMU, tactile, and fusion. Selected against the task, not against the vendor deck.
A/05Technical diligence for VCs and CVCs, and preparation of diligence-ready technical materials for robotics founders raising Series A through C.
A/06Simulation environments that reflect the real deployment site with sufficient fidelity to test operational scenarios, not just visualise the fleet.
Not exhaustive. If a scenario below sounds like you, thirty minutes on a call will tell us both whether there is a fit.
Pilots and production are different regulatory, operational, and risk regimes. The security posture that carried a pilot rarely survives contact with a production fleet.
Almost always a data problem, not a model problem. Custom training data on the specific task usually resolves it faster than another model iteration.
We prepare the technical documentation, architectural diagrams, and defensive posture that carry a raise past sophisticated diligence without becoming its own project.
Our partner network covers prototype through low-thousands under NDA with the IP posture your legal team will accept.
We help draft the answer, harden the underlying practice, and produce the audit-ready documentation the follow-up will require.
We do this diligence for operators considering their first fleet. Deliverable is a written go, no-go, or wait, with the reasoning that supports it.
Every engagement begins the same way. Discovery. Written proposal. Defined scope. Defined deliverable. We are not a body-shop and we are not a research group.
Thirty minutes. What you are building, what you are worried about, and what a good outcome looks like six months from now.
Written scope, deliverables, timeline, and pricing. Sized to the engagement, not to your budget ceiling.
Weekly progress against defined milestones. No status theatre. Handoff artefacts owned by your team from day one.
Optional ongoing advisory once the delivery closes. Fixed-fee retainer, capped hours, defined response times.
Operators deploying humanoid or industrial robotics at scale, integrators building on robotics platforms, and enterprises evaluating physical AI as a strategic capability. We work with founders, chief technology officers, heads of security, and boards.
Materially. IT cybersecurity assumes failures produce digital consequence. Robotics cybersecurity assumes failures produce physical consequence. The control set, the incident response posture, and the audit expectations are structurally different.
Yes. We capture demonstrations against defined tasks, with sensor packs calibrated to your platform, and deliver under a licensing regime that supports your downstream use. General-purpose data sets rarely produce the narrow competence a production deployment requires.
Yes. Our partner network covers prototype through low-thousands. We coordinate design-for-manufacture, partner selection, quality gates, and IP posture. Execution sits with the shop best suited to the part.
Cybersecurity readiness: two to six weeks. Architecture: six to twelve weeks. Programs: three to six months. Data capture: scoped to volume. Manufacturing: prototype through production, typically nine to eighteen months.
Always. Every scoping conversation is under an executed NDA before any material is exchanged.
Thirty minutes. Under NDA. No commitment beyond mutual honesty about whether we are the right team for the work you are trying to do.