
Humdov
Backend & ML Engineer
About the Project
Humdov is a faith-based community platform serving 3,000+ members with sermons, Bible studies, and discussion groups. I built the backend services and implemented the ML-powered content recommendation engine that personalizes each member's experience based on their engagement patterns, interests, and community activity.
What It's Made Of
Community web app
Sermons · studies · discussion groups
The member-facing product serving 3,000+ people: sermon library, Bible studies, and discussion groups, with a personalised feed rather than a reverse-chronological list. What a member sees first is the output of the recommender, which is the whole point of the platform.
Content API
FastAPI · Python
The service layer I built: content catalogue, member profiles and engagement events, group membership, and the endpoints that serve personalised feeds. Chosen for throughput on content delivery, with recommendation reads kept off the slow path so a feed request never waits on model work.
Recommendation engine
Collaborative filtering + content-based
A hybrid recommender: collaborative filtering over member engagement patterns combined with content-based matching on topics, scripture references and speaker. The hybrid exists because pure behavioural signals on spiritual content produce recommendations that feel reductive — someone who watched three grief sermons does not only want grief sermons.
Async worker
Celery · Redis
Background processing for everything the request cycle should not carry: recomputing recommendations as engagement accumulates, batch model refreshes, and content indexing. Keeping ML work on a queue is what lets the API stay fast while the model stays current.
Key Highlights
- Built recommendation engine using collaborative filtering and content-based methods
- Implemented background job processing with Celery and Redis for async ML tasks
- Designed the API layer with FastAPI for high-performance content delivery
- Content recommendation system increased user engagement by surfacing relevant sermons and studies
- Platform serves 2,500+ active community members
Technical Challenges
Building a recommendation system for spiritual content required sensitivity to context that purely algorithmic approaches miss. I combined user behavior signals with content metadata (topics, scripture references, speaker) to create recommendations that felt personally relevant without being reductive.