Position: ML Engineer

Location: Bangalore

Working Days: Monday to Friday

Experience Required: Upto 5 years

ML Engineer, Matchmaking and Recommendations

THE PROBLEM

We are looking for someone who can solve matchmaking humans. For each user we have:

- Bio data and hard preferences (age, height, religion, city, and so on)
- Photos
- In-depth conversations with our AI about who they are as a person and what they are looking for in a partner
- Feedback they have given our AI on every person we have recommended to them

The question you will own: how do we use all of this to matchmake people successfully, so that the match rate goes up, and so that a larger share of matches become happy matches. A happy match is one where the couple has a good conversation, takes it off the app, or goes on a date. This is a reciprocal recommendation problem: A recommendation only pays off if both people say yes, and a match only pays off if the conversation goes somewhere. That makes it harder than standard item recommendation, and more engaging.

WHAT YOU WILL DO

Own matchmaking end to end. We have no fixed architecture. You form hypotheses, build fast, measure honestly, keep what works, and take the winner to production. Things you will likely try:

- LLM rankers: post-train an open-weight LLM (Llama, Qwen, Gemma) to score candidates from a verbalized user context. SFT, LoRA, custom scoring heads,



reward-weighted losses, vLLM serving. Netflix's GenRec (arXiv:2608.10257) is the reference style.
- Transformer recommenders in PyTorch: SASRec / BERT4Rec-style models over feedback history, cross-encoders, sampled softmax and negative sampling.
- Two-tower retrieval with cohort features (UMAP + HDBSCAN over conversation embeddings and behavior) for cold start.
- User representations like user embeddings, etc from text, photos and tabular preferences.
- Labels and rewards: define a happy match from noisy behavior, denoise it, weight it in training.
- Reciprocal, supply-aware ranking: both sides must say yes, and good candidates are scarce.
- Offline evaluation that predicts online results, then A/B tests.

You run all of it yourself: Python, SQL on Postgres, batch pipelines, GCP (Vertex AI, GCS, Cloud SQL) or equivalent. You read papers and turn them into working systems.

THE DATA YOU WILL WORK WITH

- Structured profile and hard-preference fields
- Photos, with image embeddings and model-scored attributes
- Voice, Transcripts and LLM summaries of voice conversations with us
- Per-recommendation feedback: likes, passes, and the reasons users give us
- Outcome signals after a match: messages, calls, unmatches, off-app and date signals
- Product analytics (session and activity events)

📌 ML Engineer (HSR Layout)
🏢 PinnacleU
📍 HSR Layout