We Chose Human Intelligence Over AI — And It Worked Better Than We Hoped

At a recent AI for Good Hackathon, our team, Rooted in Resilience, set out to solve a real-world challenge brought to us by Refugee & Immigrant Transitions (RIT) — a nonprofit that supports immigrants and refugees through English language and job-readiness programs.

Their process for collecting feedback was painfully analog:

Distribute printed surveys.
Students handwrite their responses.
Volunteers manually enter data into spreadsheets.
Occasionally, mistakes creep in.
Always, it takes too much time.

So we asked the obvious question: Can AI help?

Why AI Wasn’t the Best Answer

Like most teams at the hackathon, we started with modern LLMs — GPT-4, Claude, Gemini — to interpret scanned forms. The results?

• Text fields: surprisingly accurate.

• Checkboxes: 40–60% failure rate, even with top-tier models.

• Cost: way too high for a nonprofit running on limited funds.

We realized quickly that no matter how sophisticated the AI, checkbox detection was too flaky — and definitely not budget-friendly.

---

So We Did the Unthinkable:

We Chose to Be Intelligent Instead

Instead of brute-forcing AI into a problem it wasn’t designed to solve, we designed around it.

We built SnapScan — a complete survey automation system that works without needing advanced AI for every input. Here's how it works:

---

SnapScan: Our Solution

1. Custom Form Generator

We built a tool that lets volunteers recreate RIT’s paper forms — preserving the layout and structure, but also adding two major upgrades:

• Exact box coordinate mapping (crucial for visual detection)

• A unique QR code to tag each form to the right spreadsheet

2. Scan and Detect

Once forms are filled and scanned:

• The QR code tells the system which questions and spreadsheet to sync with

• OpenCV checks for checkbox marks using pixel-perfect box mapping

• OCR handles the occasional handwritten free-text fields

3. Auto-Sync to Google Sheets

Every response is instantly logged into a shared spreadsheet — no sorting, no manual entry, no fuss.

---

A Better Demo Than AI

Here’s how we introduced it at the hackathon:

"This is the form RIT currently uses — built in Word. Here’s what it looks like in our system. Same format, but now there’s a QR code."
"Alberto — our fictional RIT volunteer — prints and distributes the forms. Once they’re filled out, he uploads all scanned forms to a Google Drive folder."
"SnapScan reads the QR code, detects the checkboxes, extracts the handwriting, and updates the spreadsheet. All automatically."

---

Why It Worked

• 100% checkbox accuracy thanks to custom layout and OpenCV

• Zero per-form AI costs — a huge win for nonprofits

• Fully automated pipeline — from scan to sheet

• Volunteer-friendly — no training needed, no tech knowledge required

---

We didn’t win the hackathon.
We didn’t use the flashiest LLM stack.
We didn’t even qualify under the "AI requirement" strictly.

But we built something that actually works — accurately, affordably, and at scale — for a nonprofit that truly needed it.

And sometimes, that’s worth more than a trophy.

---

The Team:

1. Sanath Swaroop Mulky

2. Aditi Dani

3. Manav Chandani

4. Prithvi Elancherran

5. Armin Foroughi

---

The Tech Stack:

• OpenCV (checkbox detection)

• pytesseract OCR(text extraction)

• React + Firebase (frontend & auth)

• Google Drive + Sheets API (storage & sync)

• Node.js (backend)

---

Want to Try SnapScan or Contribute?

We're exploring open-sourcing the tool for other nonprofits.
Leave a comment or DM if you're interested in collaborating or piloting it with your organization.