Divyakush Punjabi
Content Recommendation Engine — Deep learning — IIT Ropar capstone

Content Recommendation Engine — SASRec Sequential Recommender

Transformer-based sequential recommender with automated data pipelines, built as the IIT Ropar capstone project.

Model
SASRec transformer
Result
98.47% AUC-ROC
Context
IIT Ropar capstone

What it had to solve

What someone wants next depends on the order of what they did before — a recommender that treats a user as an unordered bag of preferences throws that away. Sequence models keep it, but they only pay off if the data reaching them is clean and continuous, which makes the ingestion pipeline as much of the problem as the architecture.

The decisions that shaped it

A sequential model, deliberately

SASRec is a self-attentive sequential recommender: it uses attention over a user’s interaction history so that recent and contextually relevant actions carry more weight than old ones. That choice is what makes the model answer "what next" rather than "what similar".

Trained on TensorFlow at 25M parameters

The model is implemented and trained in TensorFlow against a 25M-parameter dataset, and evaluated on a held-out split — the 98.47% AUC-ROC figure is that evaluation, not training performance.

Automated ingestion rather than a fixed dump

Automated data pipelines feed the model instead of a one-off export. A sequential recommender degrades as soon as its view of recent history goes stale, so continuous ingestion is a correctness requirement, not an operational nicety.

Containerised and served, not left in a notebook

Docker packages the training and serving path and a Node.js service exposes it to a React client, which is the difference between a model that scored well and a system someone can call.

How the pieces sit together

Client
React front end consuming recommendations
Service
Node.js API serving the model
Model
SASRec self-attentive sequential recommender in TensorFlow
Data
Automated ingestion pipelines feeding training and inference
Packaging
Docker across the training and serving path

Built into the product

Sequence-aware recommendations

Attention over interaction history rather than static similarity.

Automated ingestion

Pipelines keeping the model’s view of history current.

Evaluated, not just trained

98.47% AUC-ROC measured on a held-out split.

Deployable

Containerised and served behind an API.

What it is built out of

Model
TensorFlow, SASRec, Python
Service
Node.js, React
Platform
Docker

What it looks like

Content Recommendation Engine — The ask: what you are watching now, and what you have watched before.
The ask: what you are watching now, and what you have watched before.
Content Recommendation Engine — What the model returns from the title you are on.
What the model returns from the title you are on.
Content Recommendation Engine — The second pass — recommendations drawn from the whole history, not the last item.
The second pass — recommendations drawn from the whole history, not the last item.
Content Recommendation Engine — A title opened: rating, synopsis and the trailer, without leaving the row.
A title opened: rating, synopsis and the trailer, without leaving the row.
Content Recommendation Engine — The same sequential model pointed at music instead of film.
The same sequential model pointed at music instead of film.
Content Recommendation Engine — Song recommendations returned against a listening history.
Song recommendations returned against a listening history.

What it produced

98.47%
AUC-ROC on the evaluation split
25M
Parameter training dataset
E2E
Model through to served API