BCA Compliance AI
Rebuilding compliance operations AI for a platform processing over a million documents a month
1 Month
Product Designer
2 weeks of effort reduced to 10 min
4,000+ Active Agents
50,000+ Monthly students
A million documents a month. Ten people. One system they trust.
Edvoy’s compliance team verified every marksheet, passport and bank letter for 50,000+ students a month by hand. The brief said “automate some document checks”. What it really asked was harder: how do you put AI into a workflow where a false positive delays a visa and a false negative sends fraud to a UK university?
<10 min
review per clean case, down from 1 to 2 hours
14
fraud checks on every document
12
university and board portals checked live
~1M
documents a month, same team size
A live rebuild of the officer dashboard with sample cases. It updates on its own. Click it to filter and open cases.
02 - The 30-second version
AI assists. Humans decide. Everything else follows.
The challenge
Every application carries 15 to 25 documents. Ten officers reviewed close to a million a month, each one a judgment call. Fraud was getting sophisticated: grades edited in Photoshop with matched fonts, bank mail sent from Gmail, one agent submitting twelve students with suspiciously similar results.
What I did
Designed the full verification system: document intake, on-device OCR, a 14-check fraud engine, live portal lookups, financial document validation, AI pre-screening interviews, confidence-based routing, and a governed override path that keeps officers in charge.
My role
Defined the “AI assists, humans decide” principle and designed every workflow around it
Designed confidence routing (PASS, REVIEW, FAIL) and the officer review queue
Shaped the move from spatial regex OCR to an on-device vision-language model
Designed the override, escalation and calibration loop with compliance leadership
Built the product with Claude, from flows to working screens
The constraints
Stakes in both directions: a false positive delays a visa, a false negative ships fraud to a university
Nothing leaves the device: passports, grades and bank data can’t go to a cloud OCR API
Messy inputs: bilingual passports, hundreds of marksheet layouts, screenshots posing as originals
Trust or nothing: an accurate system officers don’t trust gets ignored
The key insight
The system’s value isn’t its accuracy rate. It’s whether officers trust it enough to act on it. So the AI flags, scores, routes and shows evidence. It never approves or rejects. That one constraint shaped every screen.
03 - Key decisions
Three calls that made it trustworthy
Decision 01
Route, don’t rule.
Confidence had to be honest. A REVIEW verdict means the system is genuinely unsure, not covering itself. Clean, high-confidence cases pass automatically. Everything else lands with an officer, including likely fails, because the machine never rejects a student on its own.
Decision 02
Read meaning, not position.
Version one used bounding boxes and regex. Bangladesh passports have Bengali and English labels; an Anna University marksheet looks nothing like a VTU transcript. The rules grew to hundreds of lines and still broke.
An on-device vision-language model replaced roughly 400 lines with one structured prompt that returns validated JSON, and no student data leaves the machine.
Decision 03
Make overriding cost a sentence.
Officers can and should disagree with the system. But overrides need a documented reason and a compliance lead’s approval. The lead then decides whether the case should teach the model or stay an exception.
04 - Context
About ten officers. A million documents. All by hand.
Edvoy processes applications for students heading to UK and EU universities from India and Bangladesh. Every application carries academic and financial documents. Before any of it reaches a university, a compliance officer verifies it. The volume was a problem; the deeper one was fraud that tired human review was less and less able to catch.
Edited grades
Marksheets retouched in Photoshop with matched fonts, and math that almost adds up.
Portal mismatches
Documents claiming results the university’s own system doesn’t show.
Fake bank mail
Personal Gmail addresses dressed up as official bank correspondence.
Agent patterns
One agent, twelve students, grade curves too similar to be coincidence.
So we asked
How do we build an AI system that takes away the repetitive verification burden without taking human authority away from the decisions that matter?
05 - Try the product
The system, rebuilt as working pieces
Each card is a real part of Compliance AI, rebuilt in code with sample data. They play on their own. Click one to use it.
The 14-check run
Pick a document and watch every check report back.
Edits the eye can’t see
Font layers and error levels reveal retouched grades.
MRZ cross-check
Catches a genuine passport with an edited photo page.
Portal routing
Three layers decide which university to ask.
Portal proof
The claimed result, checked against the source.
Bank email validation
Real bank senders pass. Gmail in a suit doesn’t.
Agent-level patterns
Fraud that only shows up across many files.
Officer review queue
Pre-triaged cases with the evidence attached.
AI pre-screening interview
Six signals, one flag, an officer decides.
06 - Outcome
Same team. A scale that would have needed three times the people.
85–90%
Less review time per clean case: from 1 to 2 hours down to under 10 minutes. The engine now covers every fraud category the team had seen, plus some they had no way to see at all. Officers moved from being the only detection layer to being the decision layer on top of one.
07 - Reflections
The hardest problem wasn’t fraud detection. It was governance.
Trust beats accuracy
An accurate system that officers don’t trust gets ignored. One that shows its evidence and defers on genuinely uncertain cases gets used, and gets better as people catch its errors.
Accountability needs a path
If every uncertain case can be overridden without a record, you’ve built a tool that looks automated but is still manual, plus the liability of an AI layer nobody audits.
Add signals, not gates
Making the AI interview a signal next to the documents, instead of a gate, kept human authority central while reaching a layer the pipeline couldn’t.