Case study

Bihar DPIP — Heritage Digitization with AI Pre-Screening

ClientArt & Culture Department, Government of Bihar

EngagementEnd-to-end design, development, and deployment

Delivery4–5 weeks (aggressive government timeline)

StackWeb platform, on-premise AI (CLIP + MobileNet), State Data Center hosting

About the client

The Art & Culture Department needed a Digital Public Interaction Platform (DPIP) to crowdsource and digitise heritage artefacts and manuscripts from citizens and institutions — following the approved heritage preservation brief and DPIP process flow.

The goal: citizens upload artefacts → AI pre-screening → expert human review → public digital archive on state infrastructure. Pure web, fully responsive on mobile browsers, English + Hindi, low-bandwidth friendly. No native app in this phase.

The challenge

High-volume citizen submissions with photos, metadata, GPS, ownership declarations, and legal consent — each needing a tracking ID and status visibility.
Manual review of every upload would overwhelm a small expert panel and slow archive growth.
Government procurement required AI that runs on-premise on State Data Center servers — no dependency on paid third-party APIs or external data egress.
Review workflow needed two expert reviewers per item, escalation paths, and separate queues for clean vs. flagged submissions.
Approved artefacts had to become publicly accessible digital archive entries with admin analytics on submission volume and stage.

What we built

Citizen portal

Registration/login, dashboard, and “Upload Artefact/Manuscript” flow.
Mandatory photos + details (type, material, language, estimated age), auto GPS capture.
Ownership selection, contribution option (donate vs. report only), legal declaration.
Instant tracking ID and real-time status tracking in the citizen dashboard.
Email notifications on status changes.

AI pre-screening (on-premise)

AI runs immediately after submission and only flags items for human review — it never auto-approves or rejects.

Duplicate detection — CLIP ViT-B/32 embeddings compared against previously approved artefacts; similarity above 92% flags as “Possible Duplicate”.
Image authenticity check — MobileNetV3-Small detects heavy editing, AI-generation, or manipulation; confidence above 75% flags as “Image Authenticity Issue”.

Runs locally on government VMs (4-core / 8 GB minimum; 8-core / 16 GB recommended). Zero recurring API cost.

Expert review workflow

Submissions routed to Normal or Flagged queue based on AI output.
Two reviewers per item with admin escalation when required.
Reviewer remarks on rejection; AI flagging reasons visible to reviewers only.

Post-approval & admin

Digital entry creation and archive storage (publicly accessible).
Admin analytics dashboard — submission counts and status by stage, real-time, exportable.

Process flow

1Citizen registers and uploads artefact with photos, metadata, GPS, and legal declaration.
2System issues a tracking ID; AI pre-screening runs on-premise.
3Item routes to Normal or Flagged expert queue.
4Two reviewers (+ escalation if needed) make the authenticity decision.
5Accepted → digital archive entry; Rejected → citizen notified with remarks.
6Citizen tracks status in dashboard at any time.

Outcomes

4–5 weeks

Delivery timeline

4–5 weeks from kickoff to UAT handover (per approved scope)

English + Hindi

Languages

Full English + Hindi citizen and admin experience

~85–90%

AI accuracy

~85–90% pre-screening accuracy on lightweight on-prem models; human reviewers provide final authority

~10–15% of submissions

Reviewer efficiency

Flagged queue isolates ~10–15% of submissions needing deeper review (duplicates + authenticity issues), letting experts focus manual effort where it matters

₹0 recurring AI API spend

Infrastructure cost

₹0 recurring AI API spend — models run on existing State Data Center VMs

100%

Citizen transparency

100% of submissions receive a tracking ID and dashboard status visibility

public digital archive

Archive growth

Approved artefacts published to a public digital archive with exportable admin analytics

Impact at a glance

AI accuracy~85–90%

Reviewer efficiency~10–15% of submissions

Citizen transparency100%

Technical notes

Uploads limited to images + PDF (no video in scope).
Hosting and maintenance on state infrastructure post-handover.
Full source code ownership and handover (Git, documentation, training) to government.

Technologies

Web platform

Responsive, low-bandwidth-optimized citizen and admin interfaces

CLIP (duplicate detection) + MobileNetV3-Small (authenticity) — open-source, on-prem

Workflow

Dual-queue review, escalation, email notifications

Hosting

State Data Center (government-operated)