Our first milestone is a repeatable deployment path: signed artifacts, per-target execution profiles, device policy, and measured rollback across i.MX95-class, Jetson Orin, and Jetson Thor. Dates are targets; evidence is what moves them to done.
Each stage names the evidence it must produce. Nothing on this site is called measured until it has been measured on our bench and validated on a customer workflow.
This is a deliberate inversion of how edge AI is usually sold. Benchmarks are quoted on a vendor's reference board, at a favourable batch size, without the thermal envelope, the memory a real application needs, or the update that eventually fails. A fleet operator cannot plan on that. We publish per execution profile: accuracy on the customer's task set, time-to-first-token and sustained throughput under thermal load, resident memory including KV cache, board-level power, and rollback time under fault injection.
The same rule governs the roadmap itself. A stage is done when its evidence exists — a signed artifact verified on the device, a measured baseline per profile, an audited escalation with its cost logged, a design partner who ran a real workload — not when a date passes. Dates on this page are targets; the evidence column is the contract.
Validate quantized small-scale inference on i.MX95-class, Jetson Orin, and Jetson Thor hardware. Establish the signed artifact format, per-target execution profiles (TensorRT on Orin and Thor; Neutron NPU on i.MX95), local policy gates, and baseline latency, power, memory, and rollback measurements.
Evidence: signed artifact verified on device; measured baseline per profile; rollback demonstrated.
Implement controlled escalation for work that exceeds the edge envelope. Evaluate provider options while preserving policy routing, context minimization, audit trails, and unit-economics measurement.
Evidence: auditable, policy-authorized escalation with minimized context; cost per escalation logged.
One narrow skill, a golden set and machine judge, a frontier teacher, an open base model under ~1B parameters distilled to the skill, quantized for i.MX95-class and a CPU-only lab target, running inside the signed artifact with judge-gated fallback. See Models.
Evidence: student beats the untuned base on the golden set across seeds; tokens/s and resident memory measured on the board; escalation path exercised.
Run an early on-prem inspection or firmware / fleet-operations pilot on real hardware, with signed model artifacts, staged rollout, audit trails, and human approval for consequential actions.
Evidence: 2–3 design-partner evaluations; one paid or committed pilot.
| Time | Product milestone | Commercial milestone |
|---|---|---|
| 0–3 months | EdgeRuntime demo; signed artifact; local verification | 25+ customer discovery interviews; ICP selected |
| 3–6 months | Staged deployment, rollback, policy gate, benchmark pack | 2–3 design-partner LOIs / evaluation agreements |
| 6–12 months | Cloud escalation, audit trail, productized control plane | First paid or committed pilot |
| 12–18 months | Repeatable deployment package | Seed-ready evidence: pilot outcomes, pipeline, early recurring revenue |
The architecture is locked around a few decisions that do not change as targets are added. Each row is a commitment, not a feature list.
| Area | Decision | Why | Status |
|---|---|---|---|
| Artifact | One signed model-artifact format carrying weights, execution-profile id, compatibility rules, policy version, evaluation record, and provenance. | Fleets mix silicon; the artifact must not. | defined |
| Execution profiles | Per-SoC profiles — imx95-npu-int4, orin-trt-int4, thor-trt-fp8; android-npu-int4 planned. CPU fallback always defined. | Same artifact, predictable behaviour on each target. | validating |
| Release | Build → evaluate → sign → verify/run. Staged rollout by percentage; previous artifact retained and verified for rollback. | Model releases inherit firmware discipline (RAUC, SWUpdate, JetPack OTA). | defined |
| Policy gate | Default-off cloud path; metadata only unless policy opens it; deterministic routing with a logged reason; append-only audit with hashes, never prompt bodies. | The model does not decide whether it may act. | defined |
| Cloud plane | Provider-aware escalation to Vertex AI, Amazon Bedrock, Microsoft Foundry; per-tenant isolation of context, keys, and retention. | Buyers choose the cloud; ByteHub owns the boundary. | pilot next |
| Evidence | Accuracy, time-to-first-token and sustained throughput under thermal load, resident memory including KV cache, board power, rollback time under fault injection — per profile. | Publish a compatibility matrix, not a benchmark slide. | Q4 2026 |
| Engineering base | Yocto / BitBake, Linux kernel, U-Boot, Device Tree, RAUC / SWUpdate, DDS. Toradex SMARC i.MX95, Jetson Orin, and Jetson Thor boards on the bench. | The layers AI products usually ignore are where releases fail. | in hand |
Silicon vendors are buying the deployment layer for their own chips. Cloud vendors are pushing agents toward the edge. OTA vendors update the OS. The governed model-release layer for mixed fleets — i.MX, Jetson, Android, x86 — is unowned. That is the position we are taking.
Land with the team responsible when a device release fails: firmware and fleet operations. Beachheads: private on-prem inspection AI and the firmware / fleet-operations copilot. 2–3 design partners, one paid pilot.
Every artifact, profile, policy, device, and evaluation result traceable per tenant. Cross-silicon (i.MX95, Orin, Thor, Android) and provider-aware (Vertex, Bedrock, Foundry) by design, so no vendor lock decides the buyer's roadmap.
Once the release path is trusted for one workload, it carries the next: perception, agents that act on equipment, escalated reasoning. Per-device runtime licences plus a per-tenant control plane compound with fleet size.
We send short, evidence-first updates to design partners, investors, and cloud partners. No marketing cadence; updates land when a milestone does.
