BergFlow · Controlled record and QA layer
A controlled workflow for every label, from intake to delivery.
BergFlow controls intake, review, QA and validation for multimodal and physical-world data. Every record is versioned, and every exception reaches a named owner.
- 01 IntakeBatch uploadTemplate, guidelines, SOP
- 02 SchemaSchema assignImage, audio, text, sensor
- 03 RouteReviewer routeNamed owner per item
- ID
- R-1048
- Schema
- image.v3
- Lineage
- 3 versions
- Owner
- Named reviewer
- Evidence
- On every edit
- 04 QAQA passSampled, then signed off
- 05 OutDeliveredLocked v3 with lineage
- What BergFlow controls
Intake, schema, routing and the governed record.
- How review works
Exceptions go to a named owner with the evidence attached.
- Modalities
Eight data modalities on one governed record.
- 96 to 98%Review accuracy, typical
- 2+QA gates per batch
- 8Data modalities on one governed record
Review accuracy is above 90% as a rule, typically 96 to 98%, agreed per project in the SLA.
How it runs
Three controls between a raw batch and a delivered one.
The same three controls apply whether the record is a product image, a call recording or a synchronised robot take.
01Ingest and queue control
Batches arrive with the task template, guidelines, ground truth and SOP attached. Queues are sized and chunked before work starts, so nothing is labelled against an undefined rule.
02Context-rich labelling
Annotators and reviewers see the item, the rule, earlier decisions and reviewer notes in one view. Disputed items carry their history to the next reviewer.
03Quality gates and validation
Every batch passes at least two QA gates: sampled checks against ground truth, then reviewer sign-off. A batch that fails its gate does not ship, and the failure stays on the record.
Modalities
Built for the data your models actually train on.
Eight modalities, one governed record per item, the same QA gates on each.
- Text curationFiltering, rewriting and ranking text to a rubric.
- Image reviewBoxes, polygons, keypoints and attributes, reviewed.
- Video segmentationFrame-accurate segments with continuity checks.
- Audio transcriptionTranscripts, speakers and intents, multilingual.
- 3D point-cloudLidar and depth labelled in 3D, synced to camera.
- Preference dataPairwise and ranked judgments for model tuning.
- Golden setsGround-truth sets that QA is measured against.
- Reviewer QASecond-pass review with reviewer accountability.
Proof
Annotation quality you can measure.
Every label in a BergFlow batch can answer two questions: who made it, and who checked it. Sampled QC runs against ground truth, reviewers sign off disputed items, and the result is written to the record before anything ships.
- Ground-truth reports per batch
- Accept and reject decisions with reviewer notes
- Skipped and expired tasks tracked as states, never lost
- Delivery exports with lineage attached
- Gate 1 · sampled QCChecked against ground truth5% sample, 97.4% agreement, pass
- Gate 2 · reviewer sign-offDisputed items resolved2 items reworked with notes, signed off
- Deliveredv3 locked with lineageExport in the agreed format
The record
BergFlow keeps every record, version and review state connected.
It is the control plane under each service. Work is versioned, exceptions go to a named owner, approvals are recorded and the audit trail stays attached to the decision.
- One governed record across image, video, audio, text and sensor data
- Exception routing to a named owner, with the evidence attached
- Approvals and an audit trail on every change
- Source
- Image, audio and text
- Route
- Annotate, then specialist review
- Owner
- Named reviewer
- Evidence
- Attached to every edit
- v1First pass by BergAuto
- v2Exception routed to reviewer
- v3Approved and audit logged
Questions
Questions about BergFlow
What BergFlow is, what it supports and how QA works.
BergFlow
Is BergFlow a self-service annotation tool?
No. BergFlow is the platform BergLabs runs managed annotation and review operations on. You see the queue, the reviewer decisions and the reports; BergLabs operates the work to an agreed SLA.
Which data types does BergFlow support?
Text, images, video, audio, 3D point clouds, preference data, golden sets and reviewer QA tasks, each held on one governed record per item.
How does QA work in BergFlow?
Every batch passes at least two QA gates: sampled checks against ground truth, then reviewer sign-off. Review accuracy is above 90% as a rule, typically 96 to 98%, agreed per project in the SLA. Failed batches do not ship, and the failure and the fix are logged on the record.
Can BergFlow work with our existing tools?
Yes. BergFlow takes batches from your storage or pipeline and returns delivery exports in the format your training stack expects, with lineage attached.
Send us one queue or a sample batch.
We grade it and come back with scope, SLA and a price, with the process on BergFlow tuned around your rules.