BergFlow
BergFlow is an invite-only annotation and review workspace. It runs annotation work from the first invite to client delivery, and every task keeps its owner, its review history and the evidence behind each decision.
Core objects
- Project
The container for one programme: the task template, guidelines, ground truth, SOP files and quality settings.
- Batch
One upload of source data: a CSV or Excel file, with an optional ZIP of media. Each row becomes a task.
- Task
One unit of annotation work, one row from the batch, with its status and full review history.
- Delivery Batch
A curated set of completed, accepted tasks, ready to export to the client.
Quality pipeline
- L1, annotation. An annotator completes the task against the template and guidelines.
- L2, review. A reviewer accepts, rejects or requests changes, with a comment. By default every L1 task is reviewed.
- L3, QC audit. A random sample of annotated and reviewed work, 5% of each by default, is audited for anything L2 missed. An L3 rejection routes the task back to the annotator with a mandatory reason.
Sampling is set per project and can be overridden for a single batch.
Roles
- Managers create projects and batches, invite the team, control project access, export results and read the reports.
- Reviewers accept, reject or request changes with a comment, and run L3 QC audits when they are assigned to QC.
- Annotators work tasks served from active batches, and can skip a task only with a comment.
Operational notes
- Uploads. CSV or Excel, with an optional ZIP of images matched to tasks by file name. Data is validated against the task template before any task is created.
- Task templates. Input and output columns, content groups, and output fields such as dropdowns, checklists, numbers and cascading selects. Templates are versioned, so existing batches keep the version they started with.
- Display and annotation rules. Optional rules that highlight or mask words and guide answers with SOP checks. They change what annotators see, never the stored data or the export.
- Batch priority. High, medium or low priority sets the order tasks are served in. An archived batch stops serving tasks until it is restored.
- Task buckets. Each annotator works from a bucket of up to 10 tasks. Leaving the workspace releases the queue so no one holds work they are not doing; submitted work is always kept.
- Skipped tasks. A skip needs a comment. Managers see who skipped a task and why, and can reopen it to the queue.
- Ground truth. GT datasets measure accuracy per batch and per output field.
- Reports. Organisation-wide user reports, project workload by person, output field summaries and GT accuracy.
- Exports. CSV by timeframe and column group, with the L1, L2 and L3 decisions and the final output for every task.
BergAuto
BergAuto is a visual workflow automation studio. Workflows are built from blocks on one canvas, run in batches or live, and every step is scored before it acts.
Blocks
- Input
Receives the work: files, records, messages and webhook events.
- AI
Model steps that classify, extract, draft or decide, each with a confidence score.
- Transform
Shapes data between steps: mapping, cleaning, splitting and joining.
- Condition
Routes a run by rule or by confidence band.
- Database
Reads and writes records in connected databases.
- API
Calls connected systems through their APIs.
- Wait
Pauses a run for a named person or an external event, then resumes.
- Output
Writes results back and records the run in the ledger.
Run surfaces
- Demo Showcase. Walkthroughs of standard workflows on sample data.
- Workflows. The governed workflows that run client work, versioned.
- Agents. Agent configurations with tools and a defined limit of authority.
- Eval. Evaluation runs that score workflows and agents against test sets before release.
How they work together
BergAuto proposes and executes; BergFlow records, routes and audits. Every BergAuto run is written to BergFlow. When a step falls below its threshold, a Wait block pauses the run and BergFlow routes the item to a named reviewer. The reviewer’s decision is versioned in BergFlow, and the run resumes.
| Responsibility | Owned by |
|---|---|
| Records, versions and lineage | BergFlow |
| Routing and exception queues | BergFlow |
| Approvals and the audit trail | BergFlow |
| Confidence scoring and thresholds | BergAuto |
| Automated steps and agent actions | BergAuto |
| API, webhook and system actions | BergAuto |
Controls to preserve
Four controls keep a deployment governed. Keep them when you integrate, extend or hand work over.
- Thresholds agreed before go-live. Trust Gate bands are set per workflow, and every change is versioned with who made it and why.
- A human wait state before irreversible actions. Payments, access changes and customer commitments pause for a named person.
- One governed record per item. Integrations write through BergFlow, never around it, so the audit trail stays whole.
- QA sampling and GT reports on every batch. Automated and human work are both sampled against the agreed standard.
Next steps
Start a pilot
Run one workflow for four weeks with these controls in place.
Integrations
Connect email, documents, voice, CRM, ERP and databases.
Security and governance
Certifications, access, audit and data boundaries.