AI · Healthcare · OCR
From Paper Form
to Smart Database
OCR, computer vision, and human oversight turn paper questionnaires into an anonymous database, queryable in natural language. Zero manual entry, zero personal data.
July 2026
Stefano Berrettini
In short
SOLVEI Survey digitizes paper satisfaction questionnaires with a hybrid pipeline: OMR for fast, deterministic optical recognition of checked boxes, a Vision AI model for ambiguous cases, smart merging of the two results, and final review always handled by a human operator before data enters the database. No personal data at all: the result is an anonymous database, queryable in natural language.
2Reading Engines
0Personal Data
NLNatural-Language Queries
100%Human Validation
The problem
Thousands of questionnaires, zero data
Italian outpatient clinics hand out tens of thousands of satisfaction questionnaires every year. Patients fill them out, staff collect them, and then they end up in a box. Transcribing every form by hand takes hours of work: it's slow, expensive, and error-prone. The data physically exists — but it's inaccessible.
The result is that healthcare facilities have no real picture of how their services are perceived. They don't know whether wait times are seen as the main problem, whether staff courtesy is rated differently across departments, or whether satisfaction has shifted over the last six months. They have the questionnaires — but not the data.
📋 The typical questionnaire
A single A4 sheet, two pages, eleven questions with checkboxes. Ratings from excellent to poor, a few demographic fields, a space for free-text comments. Simple for the patient to fill out. Complicated to process at scale without an automated system.
The pipeline
How automatic reading works
SOLVEI Survey uses two technologies in parallel to read every questionnaire, playing to each one's strengths.
01
OMR — Optical Mark Recognition
Geometrically analyzes the position of checked boxes on the sheet. Fast, precise, deterministic: if a box is checked with a clear mark, OMR sees it with no uncertainty and no need for AI.
02
Vision AI — Visual Page Analysis
A computer-vision model analyzes the questionnaire's image holistically, the way a human reader would. It resolves ambiguous cases: half-checked boxes, faint marks, different angles.
03
Smart Result Merging
The two systems are combined: wherever OMR is confident, it decides on its own. Wherever there's uncertainty, computer vision steps in. The result is more reliable than either technique alone.
04
Operator Validation
The extracted data is shown to the operator for a final review. Uncertain answers are highlighted. The operator corrects where needed, approves, and only then does the data enter the database.
Human-in-the-Loop
The human eye, in the right place
Automatic extraction is right almost all the time — but “almost all the time” isn't enough when the data is used to make decisions. SOLVEI Survey is designed around the fact that AI gets things wrong: there are badly photocopied questionnaires, non-standard markings, folded pages.
📄
Original PDF side by side
The operator sees the scanned physical questionnaire right next to the extracted values. No blind trust in the AI: the comparison is immediate and visual.
🟡
Uncertainty highlighting
Answers extracted with low confidence appear highlighted. The operator knows exactly where to look more closely — without having to re-read everything.
✅
Fast approval
One click to approve when everything's correct. Targeted edits where needed. Data enters the database only after the operator validates it.
📊
Work queue
Scanned questionnaires enter an ordered queue. Every form's status is visible: processing, ready for review, approved.
The underlying principle
AI works fast and never gets tired. Humans bring judgment and accountability. SOLVEI Survey uses AI to take the repetitive work off the operator's plate, and the operator to guarantee data quality. Neither one is enough on its own.
The database
Anonymous data, queryable in plain language
Once validated, the data enters a fully anonymous database. No names, tax ID numbers, phone numbers, or any other data that could identify a specific person is ever stored. Just the answers: satisfaction scores, the respondent's age bracket, how often they visit.
The anonymous data can be queried in natural language. No analyst needed, no SQL required. Anyone who wants to understand the situation just asks a question — and gets an answer.
💬
“What's the average satisfaction with medical staff over the last month?”
The AI queries the database, calculates the average, and answers in plain prose with the precise figure and a comparison to the previous month.
💬
“How does the wait-time rating change over the course of the year?”
It produces an automatic chart showing the monthly trend and flags the periods with the lowest ratings.
💬
“How many patients are first-time visitors, and how do they rate the welcome they received?”
It filters respondents by visit frequency and shows the distribution of welcome scores for new patients only.
💬
“What's the most frequently mentioned strength and weakness in the comments?”
It analyzes the free-text fields and identifies recurring positive and negative themes without anyone having to read every single comment.
🔒 Structural anonymity
Privacy isn't a configurable option — it's baked into the architecture. The system is designed to never collect personal data in the first place, not to anonymize it afterward. There's no data to protect, because none is ever captured.
“A questionnaire isn't a piece of paper. It's a patient's voice. It takes a tool equal to the task to actually listen to it.”
✦
Technology stack
The software powering the platform
Processing and pipeline:
FastAPI
PyMuPDF
Vision AI
MinIO
Database and monitoring:
PostgreSQL
Prometheus
Grafana
Loki
Frontend and interface:
Next.js
React
Tailwind CSS
Docker
