← All work

StatusNeo for Jio · 2021 to 2022

JioTV and JioCinema recommendations

A recommendation platform serving 470M+ users, with a 1.5% engagement lift and user profiling pipelines that run 23% faster for 30M+ daily active users.

470M+users on the platform

The problem

JioTV and JioCinema serve one of the largest audiences in streaming. At 470M+ users, a small gain in relevance reaches tens of millions of people, and a stale profile means the wrong recommendation for just as many.

Home

LiveCricket · Hindi

The Second Innings

Picked for you · you watched The Final Over

Watch now+

Because you watched Coastline

New episodeMonsoon Diaries
Chai at Nine
Top 10The Long Platform
Kites over Jaipur
Night Market

Continue watching

The Final Over
Coastline
Hindi news, 21:00
Little Rovers

My role

Software and data engineering consultant in the platform team

Results

30M+
daily active users profiled
+1.5%
engagement, team result from recency, freshness and feedback signals
23%
faster user profiling pipelines I engineered

Stack

  • Python
  • FastAPI
  • Kubernetes
  • Dask
  • Airflow
  • MongoDB
  • Cassandra

What changes

Freshness

WithoutRecommendations lag behind what is live and what a viewer just watched.

With JioTV and JioCinema recommendationsRecency, freshness and feedback in ranking. Engagement up 1.5%, a team result.

Profiling

WithoutProfiling 30M+ daily active users strains the pipelines.

With JioTV and JioCinema recommendationsDistributed profiling pipelines I engineered, 23% faster.

Shipping changes

WithoutEvery ranking change waits for a platform release.

With JioTV and JioCinema recommendationsSmall services on Kubernetes ship and scale on their own.

How it works

At a high level. A detailed walkthrough is available on a call.

  1. 01Events
  2. 02Profile
  3. 03Rank
  4. 04Serve
  5. 05Measure
  1. Decision 01

    Recency, freshness and feedback loops added as ranking signals.

    Viewers want what is on now, and the signal from their last session is the best predictor of the next. Engagement rose 1.5%.

  2. Decision 02

    Profiling moved to distributed Python pipelines, scheduled and retried by Airflow.

    Profiling runs 23% faster, so recommendations work from fresher profiles.

  3. Decision 03

    Small, independently deployable services on Kubernetes.

    Ranking changes ship without redeploying the platform, and each service scales with its own load.