Posted 1h ago
Loading this job…
Recommendation Systems Intern (Two-Tower Retrieval)
Posted 2h ago0 views
What the recommender does today
Candidates on MailerMen see a jobs-for-you rail that is, honestly, a keyword match wearing a nice card design. Your six months go into turning it into an actual recommender.
Milestones
The first two months are candidate generation: building the user-item interaction matrix from saves, applications and dwell time, then getting an implicit-feedback matrix factorisation baseline in place so there is something real to beat. Months three and four move to a two-tower retrieval model in PyTorch with embeddings for role, stack and location. The final stretch is the part most people skip, which is online measurement. You design the A/B split, agree guardrail metrics with product before launch, and read the results with us rather than being handed a verdict.
Details
You sit in our Lower Parel office three days a week with the growth and product engineers, because arguments about a recommender resolve faster in person. Written for a final-year student or recent graduate who has read about collaborative filtering and now wants to run one against live traffic, cold-start users and all. Bring curiosity about why people click the things they click.
Responsibilities
- Build the user-item interaction matrix from saves, applications and dwell time across two years of activity
- Ship an implicit-feedback matrix factorisation baseline and measure it against the current keyword rail
- Train a two-tower retrieval model in PyTorch with embeddings for role, stack and location
- Handle cold start for candidates with fewer than three interactions using profile-side features
- Design the A/B split and agree guardrail metrics with product before anything reaches users
- Read experiment results with the growth team and decide together what ships
- Keep a running note on why each ranking decision was made, for whoever inherits the system
Requirements
- Final-year student or recent graduate who has read about collaborative filtering and wants to build one
- Solid Python, plus PyTorch or a comparable framework for training embedding models
- Understands that offline and online evaluation are different things and can say why
- Comfortable with vectors, dot products and the linear algebra underneath embeddings
- Willing to work from the Lower Parel office three days a week
Skills
Benefits
- Stipend at the upper end of our internship band, paid monthly for six months
- You own a surface every candidate on the platform sees
- A seat with product and growth engineers rather than a separate intern pod
- Conversion to a full-time machine learning role for interns who ship
- Certificate and reference letter on completion