The platform

One layer across a vendor-siloed back end.

Bonders from one vendor. Molding from another. X-ray, SAM and AOI from three more. Test from a sixth. Chipira is the layer that perceives across all of them, decides once, and acts on the process.

Request a pilot Read the technology

Diagram of the four-stage Chipira loop — perceive, decide, act, verify — with a feedback arc and the back-end tool classes it connects to. — The closed loop and the tool classes it spans. Chipira architecture illustration.

Fab edgeInference runs on your premises. Latency is measured in milliseconds, not round trips.

Vendor-neutralSECS-GEM, vendor APIs and OT interfaces across the whole back end.

AuditableEvery perception, decision and write-back is logged immutably.

The shape of the thing

Not a dashboard. A system of action.

The distinctionA dashboard tells you the yield fell. A system of action changes the recipe, proves the change in a twin, writes it to the tool and records who approved it.

Most software sold into a packaging line is a viewer. It reads from the tools, renders a chart, raises an alarm, and leaves the decision — and the consequence — with a human who is already managing four other excursions.

Chipira is built the other way round. Perception exists to produce a decision; the decision exists to produce an action on the process; the action exists to produce a verified outcome that trains the next decision. Everything else is instrumentation for that loop.

That is why the platform is deployed at the fab edge rather than in a cloud console, why the connector layer writes as well as reads, why every move is checked in a twin before it executes, and why the audit trail is treated as a first-class product surface rather than a compliance afterthought.

See the agents that do the work

Architecture

Three tiers, one contract

A cloud or private-VPC control plane governs models and fleets. A fab-edge tier runs vision, orchestration and action envelopes on site. The line tier is your existing tools, unchanged.

Three-tier architecture diagram: a cloud control plane above a fab-edge Kubernetes tier above the line tier of bonders, molding, inspection and test cells. — Model registry, fleet management and evaluation gates in the control plane; vision runtime, orchestrator and action envelopes at the edge; your tools on the line. Chipira architecture illustration.

Capabilities

What the platform is made of

Perceive

Cross-modal packaging vision

Fine-tuned models for warpage and void in X-ray and CT, delamination in SAM, bump co-planarity, bridging and die cracks in AOI, and sub-micron alignment metrology — fused into one time-aligned view.

X-ray / CT volumes
Scanning acoustic microscopy
AOI image streams
Alignment and force telemetry

Decide

Agentic orchestration

Seven agents under a packaging-fab orchestrator, with idempotent steps, human-approval checkpoints and a model router that sends each step to the right model at the right cost.

Event-driven agent runtime
Graduated autonomy levels
Model-agnostic routing
Human-in-the-loop gates

Act

Closed-loop write-back

Setpoints, alignment corrections, stacking orders, routing and dispositions are written back to tools inside action envelopes your engineers define, with fail-safe stop and rollback.

SECS-GEM write-back
Bounded action envelopes
Rollback on out-of-spec
Robotic handling control

Simulate

Package-and-line twin

An Omniverse-based twin simulates bonding, stacking and molding, auto-optimising recipe and alignment to hit the target warpage, void and yield spec before the run.

Thermo-mechanical priors
Recipe auto-optimisation
Pre-run spec validation
Ramp DOE planning

Learn

The data flywheel

Every supervised correction, approval, write-back and test result becomes a time-aligned training artifact — linking perception to real process action and real yield outcome.

Per-fab defect corpus
Per-engineer craft memory
Synthetic rare-defect library
Tenant-isolated fine-tuning

Assure

Trust surface

Grounded, cited outputs; continuous evaluation gating every model and prompt change in CI; runtime guardrails; and an immutable, yield-grade audit log.

Citation and grounding checks
Golden-dataset evaluation
Schema and safety guardrails
Immutable audit trail

Non-functional

What it has to survive

A packaging line does not pause for your software.

Reliability: 99.9% uptime target, idempotent agent steps, graceful degradation, and fail-safe tool stop and robotic handling. If Chipira fails, the line falls back to its existing control — it does not stop.
Performance: High-throughput, low-latency inspection at the edge and real-time bonding and alignment control. Design target: sub-100 ms for inference decisions that affect routing or review [ASPIRATIONAL].
Security: SSO and RBAC, encryption in transit and at rest, per-tenant isolation, yield-grade audit logging, SOC 2 programme, strict process-IP protection and an on-premise option.
Observability: Full tracing, twin-based and golden-dataset evaluation gating in CI, and runtime guardrails on every agent action.
Compliance: Aligned to warpage, void and defect classification standards and package-reliability requirements where relevant to your qualification flow.

At the edge

A day in the log

The audit trail is not a report you export at quarter end. It is the running record of what every agent perceived, proposed, wrote and verified — and which engineer stood behind it.

chipira@fab-edge-04 ▸ chipira logs --agent all --tail

10:41:02  agent.place_and_bond    setpoint.write     force 18.4 N       approved k.tanaka
10:41:02  twin.package_line       validate           warpage 61 µm      within spec
10:41:07  agent.warpage_defect    classify           void · conf 0.994  escalated
10:41:09  agent.yield_ramp        route               rework             auto
10:41:14  agent.stack_tsv         sequence           kgd 99.42%         approved k.tanaka
10:41:21  agent.mold_underfill    setpoint.write     cure 142 °C        approved auto
10:41:33  agent.place_and_bond    rollback            envelope breach    held

7 actions · 2 escalations · 1 rollback · 0 unlogged

Illustrative log output. Every line is immutable and attributable.

We do not need another alarm. We need something that already fixed it and can show me why.

Yield engineerIDM back-end operations

Composite drawn from design-partner and industry conversations. Illustrative, not a customer endorsement.

Where we differ

Incumbents and copilots versus a system of action

Every incumbent owns a tool. Nobody owns the loop.

Chipira’s own competitive framing.
Dimension	Incumbents & copilots	Chipira
Scope	A single task, locked to a single tool	The full place → bond → stack → mold → inspect → test → bin loop
Loop	Setpoints held, or measurement reported	Closed perceive–decide–act–verify at the fab edge
Data	Limited to one vendor’s own telemetry	Compounding cross-fab warpage, defect and yield corpus
Integration	Read-only, shallow, or single-vendor	Deep write-back across bonders, molding, inspection and test
Outcome	Assists a human	Performs the work; the human supervises exceptions

Modules

Five products, one runtime

Land on inspection or alignment. Expand across the loop. They share one edge runtime, one connector layer and one dataset.

Trust

Autonomy is earned, not assumed

Chipira ships with the assumption that a packaging line is the most IP-sensitive and defect-sensitive environment in manufacturing. Every default is the conservative one.

Read the security note

Grounded and cited outputs Retrieval over your recipes, package specs, design rules and classification standards — with citations, so every call is traceable to a source.
Graduated autonomy Shadow → advisory → bounded autonomy, each gate opened only by measured accuracy and twin validation.
Per-tenant isolation Data, vector stores, memory and fine-tuned models are scoped per tenant. No cross-tenant leakage, by construction.
Immutable audit A yield-grade, tamper-evident record of every agent action, approval, override and rollback.

SOC 2 Type I is in progress and Type II is planned [ASPIRATIONAL].

Platform questions

Practical matters

The gating item is almost never our software — it is tool access and change control on your side. A shadow-mode deployment on a single wedge workflow is scoped in weeks, not quarters, once SECS-GEM or API access to the target tools is granted.
No. Chipira sits above your existing bonders, molding tools, inspection systems, test cells and MES. If we ever ask you to replace a tool to make our software work, we have built the wrong product.
That is the normal case. Where SECS-GEM is available we use it; where it is not, we integrate over vendor APIs or OT interfaces, and where a tool is genuinely closed we run in perception-only mode and act on the neighbouring steps.
Yes. The fab-edge tier is designed to keep operating through control-plane outages, and an air-gapped configuration is available where model updates are delivered as signed artifacts on your schedule.

Start narrow, expand relentlessly

Land one workflow. Own the loop.

A Chipira engagement begins with a single wedge workflow, a shadow-mode baseline and one signed success metric. Everything after that is expansion.

Request a pilot See how we price