0/day
emails in one inbox, five kinds of work mixed together
each dot ≈ 10 emails, coloured by queue
A documentation desk's shared mailbox
Shipping document verification · from inbox to discrepancy report
Every bill of lading checked
before the ship sails.
A documentation desk gets hundreds of emails a day: document checks, new shipping instructions, invoice queries, notices and spam. Tveris reads all of them, compares every draft bill of lading with its shipping instruction field by field, and tells you what needs a person — with the evidence.
The problem
The work is not hard. Finding it, and doing it exactly, at volume, is.
0%
of draft bills of lading differ from their instruction
63 clear46 mismatch20 unclear
Tveris test inbox · 129 SI vs BL checks
0 fields
checked by eye, each written a different way
Port of LoadingLoad PortPOL装货港
= one field
Real SI and BL layouts in the test inbox
1 in 13
emails is spam or phishing, sitting next to real orders
0 spam 480 real
Tveris test inbox · spam and phishing quarantined
Inside
Built for the desk, not for a demo.
The whole inbox on one map — all 520 emails, from the last run. Every dot is an email, grouped under the queue it belongs to.
Mismatch
46
a field differs from the SI
Needs a person
20
missing, unreadable, wrong document or blank
Verified
63
all seven fields matched the SI
In the app the map is live: pan, zoom, replay the run, and open any email down to its seven fields.
How it works
Five steps, and a person exactly where one is needed.
- 01
Read the inbox
Gmail or a mail store, with every attachment: text, Excel, Word, PDF and scans.
Read-only
- 02
Sort into five queues
Claude classifies each email; a rule engine checks the same one. Disagreements go to a person.
AI + rules
- 03
Extract seven fields
Parties, both ports, containers and gross weight, matched by meaning, not by label.
Any layout
- 04
Compare and report
Plain code judges SI against BL, so formatting never becomes a false alarm.
Deterministic
- 05
Escalate, don't guess
Missing, unreadable, wrong or blank goes to review with its evidence, and the report updates.
Human in the loop
FAQ
Questions worth asking.
Ready to see it on real mail? Open the dashboard, or connect a mailbox and let Tveris read a fortnight of it.
Where does the AI actually decide?
Claude classifies every email and explains why, reads scanned PDFs with vision, and maps layouts the parser has not seen. It never decides that two values match: the seven-field comparison is deterministic code, so the same documents always give the same verdict.
What happens when it cannot decide?
The email goes to review with the reason — missing attachment, unreadable file, wrong document type, or a blank value — and the evidence. A person confirms it or types what the document really says, the comparison runs again on their values, and the report updates.
Does it read my whole mailbox?
Gmail access is read-only and covers recent mail and its document attachments. The token is kept encrypted in your own browser, never in our database, and disconnecting removes it. Nothing is ever sent or deleted on your behalf.
Which document formats are supported?
Plain text, Excel, Word, PDF with text, and image-only scans through vision. A file that cannot be opened is reported as unreadable instead of being skipped quietly.
How do you avoid false alarms?
Values are normalised before comparison: 131,058 KG equals 131058, a repeated city-state equals the city, and a missing port code is formatting, not a discrepancy. A wrong port is still caught, because the name decides, not the code.
Can I check the numbers myself?
Yes. The discrepancy report lists every check with both values side by side, the export produces the organisers' submission file, and the whole pipeline runs from the command line with the tests.