Developers
The line speaks SECS-GEM. So do we.
Chipira integrates through the interfaces your tools already expose — SECS-GEM, OPC-UA, vendor APIs, MES and PLM — and exposes a small, boring, well-documented surface of its own.
Integration philosophy
We adapt to your line, not the other way round
An advanced-packaging line is a decade of accumulated integration decisions. No fab is going to rebuild it for a vendor, and any product that requires them to has already lost.
So Chipira reads what your tools already emit and writes only through interfaces they already accept. The connector layer normalises the differences; the agents never see a vendor-specific quirk; and every write goes through a single, audited, envelope-checked path regardless of which tool is on the other end.
The consequence is that integration effort concentrates in one place — the connector — where it can be specified, tested, certified and reused, instead of spreading through agent logic where it would be invisible and unmaintainable.
- SECS-GEM / HSMS for bonders, molding, inspection and test tools
- OPC-UA and vendor REST or gRPC APIs where available
- MES, YMS and PLM integration for context and write-back
- File and image drop for legacy inspection systems
Surfaces
What you can build against
Agent API
Query agent state, submit a subject for evaluation, retrieve a grounded proposal with its citations and twin verdict, and approve or reject it programmatically.
Event stream
A durable stream of perceptions, proposals, approvals, write-backs, exceptions and verifications — the same events the console renders and the audit trail records.
Envelope API
Read and version action envelopes. Widening one requires an approval flow; the API enforces the same governance the console does.
Twin API
Submit a candidate recipe or sequence for simulation and receive a predicted warpage, stress, void or yield outcome with a confidence band.
Connector SDK
Build a connector for a tool we have not yet covered, against a documented interface contract that specifies telemetry, actions and fail-safe behaviour.
Webhooks
Push exceptions, escalations, autonomy revocations and pilot-metric snapshots into the systems your engineers already watch.
Agent API
A proposal, with its receipts
Every agent response carries what it saw, what it inferred, what it wants to do, what it is allowed to do, and what the twin predicts will happen. There is no response shape in which those are optional.
{
"subject": { "package": "PKG-4471-A", "step": "post-reflow-xray" },
"perception": {
"model": "void-xray-ct@4.2.1",
"findings": [
{ "class": "void", "site": "bump C7", "confidence": 0.994 }
]
},
"grounding": [
{ "source": "spec/pkg-4471/rev-C", "locator": "§4.2" },
{ "source": "standard/void-classification", "locator": "class-2" }
],
"twin": { "verdict": "out_of_spec", "predicted_yield_delta": -0.031 },
"proposal": { "type": "route", "to": "rework" },
"envelope": { "id": "inspect-line4@7", "within_bounds": true },
"autonomy": { "level": "L2", "requires_approval": true }
}Illustrative response. Field names are indicative of the shape, not a frozen contract.
Fab-edge runtime
Installed like infrastructure, not like an app
The edge runtime is a lightweight Kubernetes deployment on fab-edge hardware. It is installed by your team, pinned to a model release you choose, and never auto-updates on a production line.
$ chipira edge install --site kh-04 --line 4
✓ hardware check GPU 2× · 128 GB · NVMe 4 TB
✓ cluster k3s v1.31 · 3 nodes
✓ runtime chipira-edge 0.9.2 (pinned)
✓ models void-xray-ct@4.2.1 · bond-align@2.8.0
✓ connectors secs-gem ×6 · opc-ua ×2 · mes
✓ envelope inspect-line4@7 (shadow)
$ chipira agents status --line 4
warpage_defect shadow acc 0.981 n=12,441
place_and_bond shadow acc 0.964 n=8,902
test_and_bin offline —
$ chipira audit tail --follow
10:41:07 warpage_defect void bump C7 0.994 → route:rework (escalated)Illustrative CLI session.
Integration surface
What we need from your line
Stated up front so your IT/OT team can assess it before a meeting.
- Tool telemetry
- SECS-GEM / HSMS event and variable access on bonders, molding, inspection and test tools; OPC-UA or vendor API where SECS-GEM is not exposed.
- Inspection data
- X-ray, CT, SAM and AOI image or volume access at line rate, by stream or by monitored drop location.
- MES context
- Lot, package, recipe, route and disposition context — read access is enough to start; write-back is what closes the loop.
- Network
- A segmented OT network path between the fab-edge cluster and the in-scope tools. No inbound internet access is required at the edge.
- Hardware
- Fab-edge compute sized to sensor count and resolution. We specify it during scoping; you procure and own it.
- Identity
- SSO integration and a role mapping that separates viewers, approvers and envelope owners.
Path to production
From access request to closed loop
-
Read-only connection
Connectors are configured read-only. Nothing can be written to any tool, by construction, at this stage.
-
Shadow mode
Agents run continuously and produce proposals nobody executes. You measure them against your engineers on your packages.
-
Envelope definition
Your process owners author the action envelope: which parameters, which ranges, which tools, which conditions, which fail-safe.
-
Assist mode
Proposals become one-click approvals in the console or through the API. Every approval and rejection trains the next model.
-
Bounded autonomy
Moves that have earned their accuracy gate execute inside the envelope, twin-validated, fully audited, instantly revocable.
OT infrastructure leadHigh-volume packagingThe integration question I care about is not how you read the tool. It is what happens to my line when your service dies mid-write.
Composite drawn from design-partner and industry conversations. Illustrative, not a customer endorsement.
Failure behaviour
What happens when something breaks
The edge loses the control plane
Nothing changes on the line. The edge runtime holds its pinned models and its current envelope and continues operating; the control plane is for governance, not for inference.
The edge itself fails
Tools revert to your existing control. Chipira is additive to the line, never load-bearing for it; a total Chipira outage degrades you to exactly the process you run today.
A write is interrupted
Actions are idempotent and transactional. An interrupted write is either completed or rolled back, and either outcome is recorded in the audit trail with its terminal state.
A model drifts
Rolling accuracy monitoring withdraws the autonomy level automatically, escalates to your engineers, and pins the previous model release pending review.
Technical questions
What engineers ask
-
No inbound access, and outbound only for model and fleet governance. Air-gapped deployments have no outbound access either; models arrive as signed offline artifacts.
-
Perception runs on fab-edge GPUs precisely so that it can meet line-rate requirements on the wedge workflow. Achievable latency depends on sensor resolution and volume and is characterised during scoping, not promised generically.
-
On enterprise agreements, yes — custom architectures and customer-supplied models can be deployed through the same versioned, gated release path as ours.
-
Every model release is immutable and semantically versioned, pinned per site, rolled out in stages, and referenced by version in every audit record it produces.
Get started
Read the docs. Then send us your tool list.
The fastest route to a real answer is your tool inventory and one workflow you would like to see run in shadow mode.