Rapta

Named devices on the line are under Rapta's read and write control. The Supercoach API covers SuperPod robotic inspection stations (connect, home, move to position, queue and cancel motion sequences, emergency stop), connected torque wrenches (update torque configuration, cancel the current tightening operation), and Basler cameras and autofocus; knowledge base installation articles cover RFID and Zebra barcode scanners. The digital I/O article documents "output signals" generated on GPIO pins to the line's automation. The Enterprise Integrations page places MES and ERP connection on the API.

MES and ERP connectors are customer-built on the API rather than native, and credential scoping and custom tools are not documented.

Sourcerapta.ai/api-documentation (Supercoach API 1.16.22), rapta.ai/knowledge-base/digital-i-o-signaling-and-connections, rapta.ai/ai-platform/enterprise-integrations; readread 2026-09-15

Rapta programs each assembly as a sequence of steps that mixes deterministic actions with AI checks, with versioning and reuse. An assembly is a programmed "sequence of steps", each a distinct action or QA decision point (Getting Started with Training). Vision QA checks run alongside torque steps, OCR capture steps, video instruction actions, data capture actions and digital I/O outputs, and the API queues SuperPod motion sequences step by step.

Branching on outcome is documented at the step: a supervisor-authorized bypass records the step as Passing and a Next bypass as Failing, and incomplete records can be resumed or pulled for rework. Versioning: "editing with visual revision control", a "revision lineage (inheritance tree)" endpoint, and roll back or roll forward of approved releases. Reuse: a Cloning Assemblies article.

Not described: first-class retry and fallback paths beyond rework.

Sourcerapta.ai/api-documentation, rapta.ai/knowledge-base/using-the-platform-training (20 May 2025), /knowledge-base/release-control; readread 2026-09-15

Checks are grounded in the customer's own parts, examples and assembly steps, but that knowledge enters by training a model rather than through a retrieval layer. The How it works page says the platform "learns your proprietary assembly techniques from just a few correct and incorrect examples" and "we automatically generate the AI training set", with a physics engine producing synthetic images of correct and defective parts. The API documents "create a new model for a specific assembly" and "initiate training of the model".

New customer knowledge cannot be added without retraining: a new assembly or revision enters by training a vision model, compiled into weights. No retrieval structure over the buyer's own content that stays queryable between runs is described. Video work instructions are authored content shown to operators, not retrieval by the agent, and OCR capture of serial and batch numbers is data capture.

Sourcerapta.ai/ai-platform/how-it-works, rapta.ai/api-documentation; readread 2026-09-15

Per-step checkpoints, escalation and pause controls are built in, configured in Rapta's Setup tabs. Supervisor lockout (release 7.4.0): the Supervisor and Next buttons that bypass an AI-checked step can be "password protected", applied to "any step type that you wish". A supervisor-authorized bypass records the step as Passing and a Next bypass as Failing, so an override is explicit. Escalation: "email supervisor when operator is stuck". Pause: the knowledge base carries a Pause Live Operation article. The checkpoint is configured per step type.

Routing approvals into ticketing tools, or explaining why a checkpoint fired, is not documented. Supervisor sign-off on new assembly revisions happens separately, under Release Control.

Sourcerapta.ai/knowledge-base/supervisor-lockout-email-alerting (5 November 2024), /knowledge-base/pause-live-operation (title in sitemap); readread 2026-09-15

Access controls are documented, but compliance rests on adherence to standards, not a held certification. Customer data is stored "encrypted at rest with RBAC", the platform applies "role-based access control (RBAC), strong authentication mechanisms, and regular access reviews", and a subset of telemetry is "visible to team admins in the team log". The knowledge base documents supervisor-level authentication for training and a Security tab where a supervisor sets a password. No SSO, SAML or SCIM is documented. Customer data stays on-premises and is never used to train other customers' models.

On compliance, the policy says "we adhere to the following ISO standards": ISO 27001, 27002 and 27018, and the ITAR page describes "ITAR-compliant practices" with products classified EAR99. No certificate, auditor, report or date is named, and no badge appears on the page. Adherence to a standard is not an attestation, and no ISO 27001 certificate or SOC 2 report is published. EAR99 is an export classification, not a security attestation.

Sourcerapta.ai/security, rapta.ai/itar, rapta.ai/knowledge-base/supervisor-lockout-email-alerting; readread 2026-09-15

Each Rapta assembly run produces a per-step record of the system's QA verdicts that a buyer can inspect after the fact. The Manufacturing Co-Pilot page describes an "assembly report with photographic traceability" integrating serial, part, date-code and batch tracking, with "each component's journey... visually recorded" and reports "archived for long term storage" for NTSB, FDA, DOE and NHTSA requirements.

The API exposes the record: assembly records per assembly, the recording for each record, "generate a QA report" and "retrieve a QA report" for a specific record, plus scheduled report delivery. Supervisor bypasses are recorded as Passing or Failing on the step, and under the Security policy a subset of telemetry and logs is "visible to team admins in the team log".

Operator metrics and Operator Analytics measure the human worker's performance, not the agent's. The model inputs and confidence behind each verdict are not documented, and neither is SIEM export.

Sourcerapta.ai/ai-platform/ai-manufacturing-co-pilot, rapta.ai/api-documentation, rapta.ai/security; readread 2026-09-15

What persists between runs is trained models and production records, not agent memory. Learned assembly models are trained weights, and work instructions, assembly records and operator metrics are the application's records. Learning across sessions is not a memory layer, and neither is the application's own database of business records.

The API's "resume an incomplete assembly record" and "get assembly records for rework" are the closest thing to run state, and they resume a production record, not a memory the agent keeps; deleting that state deletes the traceability record. No session, conversation, workflow or long-term memory is documented, so there is nothing to review, edit, delete or scope.

Sourcerapta.ai/api-documentation, rapta.ai/ai-platform/how-it-works; readread 2026-09-15

Everything runs on an on-premises AI server, and customer data stays in the customer's facility. The Security page states: "Customer data is securely stored and processed in an on-premises AI server. It never goes to the public cloud. Unlike cloud-based solutions, our customer data never leaves your facility."

The knowledge base documents the station hardware (Siemens IPC520a industrial PCs and a per-workstation Rapta Gateway with a static IP), and the API exposes the gateway's network settings. Data and processing are pinned to the customer's own site, so residency is answered by location rather than by a region list. The ITAR page adds that software, models, source code and training data are developed and kept on U.S.-based systems by U.S. persons.

There is one mode only, so no feature differences across modes arise; who patches and upgrades on-site servers is not documented; and remote support access is optional under the Security policy.

Sourcerapta.ai/security, rapta.ai/itar, rapta.ai/api-documentation; readread 2026-09-15

Industry application solutions and worked examples are named, but every assembly is learned from the customer's own examples before it runs, so nothing documented is adopted ready-made.

The How it works page lists named application solutions (Drones, Industrial Electronics, Laboratory Instruments, Precision Optics, Semiconductor, Wiring & Panel Validation); the sitemap carries application pages for panel and wire inspection, kitting and part counting, multi-spectrum FOD inspection and TorqueGuardian; and the knowledge base carries worked examples (realtime QA for electronics, bracket assembly and RMA; panel wiring validation) plus a Cloning Assemblies article.

These are industry starting points. Inspection routines that load per assembly type and work instruction templates are the customer's own trained assemblies, not packaged assets.

The application solutions are industry framings of one configurable platform, and no asset is documented that a buyer adopts and runs as is; no prebuilt, ready to run inspection model is published.

Sourcerapta.ai/ai-platform/how-it-works, rapta.ai sitemap, rapta.ai/knowledge-base index; readread 2026-09-15

Work reaches Rapta without a person starting it, through a documented machine signal. The knowledge base article on digital I/O says "input signals used to start assemblies" should follow the documented timing, and "Supercoach is looking for a rising edge" on a GPIO input from line automation (Siemens IPC520a wiring documented): a machine signal starts an assembly run with no person asking.

The inline QA checks each step from the camera as work happens rather than on request. Reports can be scheduled through the API ("change how frequently a report is emailed, and who receives one"), and supervisor lockout can "email supervisor when operator is stuck". Digital I/O, barcode and RFID scanners and email are first-class channels.

Webhooks, messaging tool or MES event subscriptions, and duplicate or race handling are not documented.

Sourcerapta.ai/knowledge-base/digital-i-o-signaling-and-connections (21 October 2023), /knowledge-base/supervisor-lockout-email-alerting, rapta.ai/api-documentation; readread 2026-09-15

Model choice in Rapta is vendor-controlled, and no provider is named. The Security page says "our AI models are trained exclusively on Rapta owned datasets"; the API trains a model per assembly and in a parts library; and the How it works page describes a "prompt based physics engine" that generates training images from text prompts.

No foundation model, provider or version is named for the vision models or the physics engine, no routing across models is disclosed, and no customer selection, bring-your-own-model or key path exists, including in the 224-path API, whose model endpoints create and train Rapta's models rather than choose among them. Training a model on the customer's examples is customization, not model selection.

Sourcerapta.ai/security, rapta.ai/ai-platform/how-it-works, rapta.ai/api-documentation; readread 2026-09-15

Rapta publishes a documented API for its own platform. The sitemap lists /api-documentation/, and the Enterprise Integrations page says "see our API documentation for more details".

That page renders a Swagger UI over an inline OpenAPI specification titled AI Supercoach, version 1.16.22, with 224 documented paths: read-write endpoints for assemblies, steps, revisions and revision lineage, assembly records and QA reports, operators and operator metrics, model training (initiate, schedule, cancel), cameras, OCR capture variables, torque configuration, SuperPod motion control, and websocket channels for coaching updates, training status and device streams. On its own documentation the platform is API-first, and the knowledge base carries an Example API Usage article.

No SDK, MCP surface, or versioning and deprecation policy is documented. Authentication is shown as operator credential login.

Sourcerapta.ai/api-documentation (Supercoach API 1.16.22), rapta.ai/ai-platform/enterprise-integrations; readread 2026-09-15

Assembly changes wait for supervisor sign-off before they reach the floor. The Release Control article says "new revisions and changes are held for digital supervisor sign-off before updates are pushed to the factory floor", unapproved assemblies are unpublished and invisible to operators, and it "works seamlessly with revision control enabling instant roll back or roll forward of an approved release". That is a quality gate in the release path of an assembly revision.

The gate is a person's approval, with no test run, score or pass rate to judge the revision against. Testing with fixtures or datasets before production, and scoring output quality over time, are not documented for the vision models. Generating training sets automatically is how the model is built, not evaluation, and operator metrics measure the human operator, not the agent.

Sourcerapta.ai/knowledge-base/release-control (22 May 2024), rapta.ai/ai-platform/how-it-works, rapta.ai/api-documentation; readread 2026-09-15

Cameras watch the physical work, and documented device interfaces control physical devices (SuperPod motion, torque wrenches, GPIO); none of that operates a graphical interface on the customer's computers, and camera based inspection has no screen to automate. The knowledge base's Screen Recording on the Rapta Station article records the station's own display for support, and machine vision of parts is the product's quality function, not computer use.

Sourcerapta.ai/ai-platform/how-it-works, rapta.ai/api-documentation; readread 2026-09-15