Dashboard
Operational overview from the database — throughput, approval rates, and per-client totals.
Rows missing a title, name, gender, address, post code, DOB, NI number, or work start date are flagged automatically — fix a cell inline, tick the rows you're happy with, then approve them all at once.
Rows where hours couldn't be confidently computed, or where hours/rate were auto-computed from an in-file rate legend, are highlighted below — fix a cell inline, tick the rows you're happy with, then approve them all at once.
| S.No | Employee Code | Employee Name | Client | Gross Hours | Rate | Gross Pay | Status |
|---|
| S.No | Employee Code | Employee Name | Client | Gross Hours | Rate | Gross Pay | Status |
|---|
| S.No | Employee Code | Employee Name | Client | Gross Hours | Rate | Gross Pay | Status |
|---|
| S.No | Employee Code | Employee Name | Client | Gross Hours | Rate | Gross Pay | Status |
|---|
What goes in
Approved when
File format
New starter batches
Every pack the agent has read, newest first. Tick batches to build a custom download, or use the arrow on a batch to take just that one.
| Label | Status | Date Processed | Missing fields | Updated |
|---|
What goes in
Approved when
Which clients
File format
Timesheet batches
Every timesheet upload, newest first, with its hours and gross pay. Tick batches to build a custom download, or use the arrow on a batch to take just that one.
| Label | Status | Date Processed | Missing fields | Updated |
|---|
Latest reconciliation
Agents
The agent that runs this workspace.
Business metrics
From the database, not just this session.
Throughput & status
Clients attempted per day, and how each import run finished.
Clients attempted per day
Last 7 days
Status breakdown
Client creation outcome
Extras breakdown
Best-effort actions attempted after each client is created — Bank Account, Customer, Supplier, Quote.
| Functionality | Attempted | Succeeded | Success rate |
|---|
Recent activity
Most recent import attempts, newest first.
| Client | Status | Extras | Updated |
|---|
Audit log
Every action on this run, in order. Append-only. Click to open.
Audit log
Every action on this run, in order. Append-only. Click to open.
+ New Agent
Finishing this wizard adds a draft blueprint to Agent Studio and to the Dashboard's Draft agents section, where you can lay out and inspect its flow. It won't run for real until a backend is wired up for it.
One file, one worker. Each PDF, photo or ZIP you pick is one person.
Queue
Intake
Extract
Decide
Filed
One click, one batch. Everything you upload together is grouped as a single batch in HITL and History.
Queue
Intake
Understand & compute
Decide
Filed
Agent line
Runs top to bottom+ New Agent
Finishing this wizard adds a draft blueprint to the Agent page and to the Dashboard's Draft agents section, where you can lay out and inspect its flow. It won't run for real until a backend is wired up for it.
Feed it a statement. It settles the books.
Every bank line matched against invoices, receipts and purchases. Certain ones are reconciled, the rest are queued for a person.
You give itOne bank statement, CSV or Excel
You get backMatches, confidence, and an exception list
It fetchesLive from Capium for Demo Company
Confidence 95/80%
Tune from your own results. These aren't accounting rules.
Pick an account and drop a statement to start.
Pick the account this statement belongs to.
- Upload takenReading the statement
- ProcessingNormalising each transaction
- InvoicesReading Capium's sales ledger
- PurchasesReading Capium's purchase ledger
- JournalsReading Capium's journals
- LLM matchingJudging each candidate
- ReconciledSettling what's certain
- Cash codingCoding what's left to accounts
—
| Run ID | Client | Steps | Status | Duration |
|---|---|---|---|---|
| run_8f3a2d | Fenwick Roasters Ltd | 5 / 5 | Success | 38s |
| run_7c3e1b | Marcus Ojo Photography | 5 / 5 | Success | 31s |
| run_6d1a9f | Ashcroft & Bell Surveyors | 1 / 5 | Failed | 21s |
Capium Login
The Bank, Sales, Purchase and Journal agents sign in to Capium with this. Saved encrypted in this app's own database — change it here any time.
Loading…
No model in this pipeline
Capium's automation is pure browser automation — it drives real screens directly and doesn't call the VLM or any other model. Nothing to configure here.
Agents
Two independent agents run this system — each its own pipeline, calling the shared VLM to make judgment calls a fixed rule table couldn't.
Draft agents
Created in the wizard, laid out in Agent Studio — not yet wired to a backend, so no run stats exist for these.
Business metrics
From the database, not just this session.
Throughput & status
Items processed per day, and where each pipeline stands.
Items processed per day
Last 7 days
Status breakdown
New Starters vs Timesheets
Recent activity
Most recent items from the database, newest first.
| Type | Worker | Status | Updated |
|---|
Per-client totals
Approved timesheet rows only — figures are final once approved.
| Client | Rows | Hours | Gross Pay |
|---|
Observability
Real counts derived from stored items. This app doesn't instrument per-call latency or token usage today, so those aren't shown here rather than made up.
Error breakdown
Items whose notes indicate the agent couldn't confidently extract something.
| Reason | Count |
|---|
Live rates
Auto-approval / review / rejection rates, computed continuously from the real database — a live proxy for extraction confidence, not per-field accuracy (that needs a verified ground-truth answer key, which only ever existed for the one-time benchmark below).
New Starter Agent
Timesheet Agent
Historical validation benchmark
A one-time offline run against a real, known-correct dataset (w/e 30.08.2026) — a dated reference point, not live production accuracy, and not recomputed automatically.
| Client | Rows checked | Exact matches | Match rate |
|---|---|---|---|
| Banham Poultry | 613 | 611 | 99.7% |
| Poundland | 174 | 174 | 100% |
| ABP Penston | verified against a source screenshot of the file's own rate legend, not row-by-row ground truth | — | — |
Each row is one upload batch — the same grouping HITL already uses — with a status derived from its own items' real statuses, not a separately tracked run concept.
| Source | Type | Items | Status | Started | Updated |
|---|
New Starter Agent
Reads a worker's onboarding documents and extracts payroll/identity fields.
Timesheet Agent
Reads a client's raw timesheet, works out its pay structure, and computes gross pay per worker.
This system uses one extraction backend — a VLM — not a set of interchangeable providers.
VLM
Shared, already-running service on this box — used for both vision extraction and text-only structure reasoning.
| Model | VLM |
| Capabilities | Vision extraction + text reasoning |
| Max new tokens | — |
| Timeout | — |
HITL behaviour
Real, persisted toggles — changes apply from here on (already-decided items aren't retroactively changed).
Timesheet category labels
Some clients' pay-category tags aren't self-explanatory on their own (e.g. a weekend premium tagged "WP"). Add a client + tag pair here and its S.No will show your label instead — e.g. client "AB Produce Packhouse", tag "WP", label "W Bonus" turns S.No "5WP" into "5 W Bonus". Unconfigured tags are left exactly as they are.
| Client | Tag / Code | Label | Enabled |
|---|