Finance ops
Three-way match & exceptions
Invoice, PO and goods receipt reconciled; short supply raises the debit note and holds payment pending approval.
Each one is useful alone. They are worth considerably more together, because the output of one is the input of the next: perception feeds the structuring, the structuring feeds the agent, and the agent's decisions become next quarter's training data.
Most "AI automation" stops at a recommendation and leaves a human to do the typing. Ours are given real tools, a written contract of what they may and may not touch, and an approval gate wherever the stakes justify one.
The contract is the product. It is reviewable by your risk team, versioned in git, and enforced by the runtime rather than by a prompt.
agent: ap.exception.resolver tools: - erp.grn.lookup # read - vision.query # read - ledger.debit_note # write - mail.send # write may_write: [debit_note, vendor_email] may_not: [release_payment, edit_vendor_master] gates: - when: value > 100000 require: named_approver sla: 4h - when: confidence < 0.92 require: human_review on_failure: route_to_queue evidence: [page_crop, video_clip] audit: immutable · 7y
Invoice, PO and goods receipt reconciled; short supply raises the debit note and holds payment pending approval.
Stop detected, clip pulled, ticket opened, maintenance paged, root-cause draft written before anyone arrives.
Claims, KYC and pre-authorisation packets assembled, cross-checked, and returned with a list of what is missing.
A general-purpose chatbot pointed at your business is a liability: it is confident, it is fluent, and it is occasionally wrong in ways nobody catches. Ours is built the other way round - it cannot make a claim it cannot cite, and when the evidence is not there it says so and stops.
Shift handovers, downtime narratives, incident summaries and audit responses - drafted from the evidence, edited by a human, never sent unattended.
Plain-language explanations of a detection, a score or a flagged exception, with the crop, the clip or the row that drove it.
Operator notes in one language, procedures in another, and the shorthand that only makes sense on your site.
Detection tells you something is there. Recognition tells you which one it is. What operations actually needs is the third thing: how many, how big, how fast, and is it wrong. We build all three, on your imagery, running on hardware that fits next to the line.
Find every instance and where it sits, at line speed, with stable identity across frames.
Which variant, which grade, which serial, which person is authorised to be in this zone.
Dimensions, counts, fill levels, coverage, and the defects nobody wrote a rule for.
Scratches, porosity, missing fasteners, wrong-part-fitted, label mismatch.
Flag, never decide. Always shows the region and the confidence behind it.
Stand counts, disease severity, weed maps, ripeness and grading.
Helmet, vest, guard-door and restricted-zone events, privacy-preserving by default.
A camera produces 86,400 seconds a day and a human reviews roughly none of it. We convert continuous footage into a timeline of events - actions, states, durations, exceptions - that you can filter, query in plain language, and attach to a work order as evidence.
Natural-language search across months of footage. Answers come back as clips, not timestamps you have to go find.
Every process step timed automatically. Standards drift; now you can see exactly when and where.
Models learn what precedes a stop, a jam or a defect cluster, and raise the flag while there is still time to act.
After a year of structured extractions and detections, you own something you did not have before: a clean, labelled, longitudinal record of your own operation. That is where the genuinely valuable models come from - the ones nobody else can build because nobody else has the data.
Near-duplicate invoices, altered totals, vendors sharing bank details, spend that breaks pattern.
Where the next defect cluster and the next unplanned stop are most likely to originate.
Credit, claims and case-priority scoring built on your own history, with reasons attached.
Distributions watched continuously; when the world shifts, you hear it from us first.
Document context understanding, OCR and handwriting, classification, splitting and grounded retrieval over archives. It is the deepest part of what we build, so it lives separately rather than squeezed into a card here.
A partial list, with the honest version of what each one takes to get into production.
| Use case | Sector | Capabilities used | Typical time to pilot |
|---|---|---|---|
| Surface & weld defect inspection | Manufacturing | vision · edge · ml | 3-5 weeks |
| Assembly-step verification | Manufacturing | vision · video | 4-6 weeks |
| Downtime root-cause & triage | Manufacturing | video · agents | 4-6 weeks |
| Invoice & three-way match automation | Cross-sector | documents · ocr · agents | 2-4 weeks |
| Clinical document abstraction | Healthcare | documents · ocr | 4-8 weeks |
| Imaging second reader | Healthcare | vision · ml | 8-14 weeks |
| Crop scouting & disease severity | Agriculture | vision · ml | 1 season |
| Weed maps for variable-rate spraying | Agriculture | vision · edge | 4-8 weeks |
| Pack-house grading & sorting | Agriculture | vision · edge | 5-8 weeks |
| KYC & loan file assembly | Finance | documents · ocr · agents | 3-6 weeks |
| Legacy archive digitisation & search | Public sector | ocr · documents | 6-12 weeks |
| PPE & restricted-zone compliance | Cross-sector | vision · video · edge | 3-5 weeks |
The fastest way to know whether this works on your data is to try it on your data. Two weeks, your infrastructure, a real prototype and a straight answer.