Applied AI in Production: From Computer Vision QC to LLM Fraud Detection
A look at how we ship AI features that hold up in production — defect detection at warehouse intake and AI-assisted identity deduplication for welfare programs.
It's easy to demo AI. It's hard to ship AI that a warehouse manager or a government auditor will rely on every day. Across projects like our AI warehouse quality-certification MVP and the PPP family-duplication prevention engine for the Government of Haryana, we've learned that production AI is mostly about the engineering around the model.
Computer vision at warehouse intake
For agricultural warehousing, we built a computer-vision pipeline that runs at the moment stock arrives:
- Product verification — catch mismatched stock at intake before it's accepted
- Defect detection — camera-based grading of commodity quality
- Automatic certification — generate a digital quality certificate from the grade
The dashboards then aggregate grading and stock across multiple warehouses in real time. The model matters, but the value comes from wiring it into the intake workflow so a bad batch is flagged before it's booked.
Deterministic first, AI second
The PPP system had to guarantee that each citizen's Aadhaar links to exactly one active family unit. That's a fraud-prevention problem where mistakes have real consequences, so we layered it deliberately:
- A deterministic dedup/rule engine handles the clear cases
- AI/ML fuzzy matching resolves the ambiguous ones — name variants, transposed fields, document mismatches
- LLM-assisted document verification and case summarization support human reviewers
AI should widen the net for ambiguous cases, not replace the deterministic rules that already work.
Making it usable
We added voice-enabled, multilingual search and an automated scheme-eligibility and notification module, so the intelligence reaches citizens and caseworkers in the language and channel they actually use.
The pattern we reuse
- Put AI inside the workflow, at the decision point
- Keep deterministic logic as the backbone; use AI for the hard middle
- Always leave a human-reviewable trail — summaries, scores, certificates
That's how AI features survive contact with production.
Have a project like this?
We build enterprise, government-scale and AI-integrated platforms end-to-end.
Start a project
