HE-Universe Multi-Domain AI Model Store

Problem

As the number of client programs grew (BAT, Unilever, OCR, face verification, speech), running each as a fully separate service became an operational burden — duplicated boilerplate, scattered deployments, and no single place to reason about the whole model estate.

Approach

• 01Consolidated BAT, Unilever (UBL), face detection/recognition, Bangla & 'MT House' OCR, material detection, beverage-industry detection, hotspot calculation, eKYC, and five NLP/conversational modules (general NLP, Zerocal, Maya, conversational guidance, speak-metric merge) behind one FastAPI app (Universe_API.py).
• 02Kept each domain in its own class/module boundary (modules/<name>/<name>_main.py) while sharing a common tools/model-handler layer for loading and serving models.
• 03Deployed as a single Hugging Face Space so every client-facing capability the team ships can be reached from one live endpoint instead of a dozen scattered ones.

Results & Impact

• One live Hugging Face Space now fronts 15+ previously-separate production modules spanning retail CV, OCR, identity, and conversational NLP.
• Gives the team a single place to add, version, and monitor new AI capabilities instead of standing up a new service per client.— CV-sourced