Back to work
AI / MLPrototype

RetailMind

Graph-based product recommendation experiment with a reproducible delivery pipeline.

Type
AI / ML
Client
Independent ML research prototype
Role
ML engineering · MLOps
Year
2024

Evidence / status

Prototype. Reported outcomes await verification.

Challenge

Recommendation infrastructure without an ML team

The work explored whether a smaller retail dataset could support graph-based recommendations and a repeatable path from training to API serving.

Constraints

  • Experiments needed reproducible data and parameters.
  • Model comparison needed more structure than notebook output.
  • The serving interface had to remain simple to integrate.
  • Retraining needed a documented path rather than manual steps.

Approach

Key decisions

  1. Graph-based collaborative signals

    LightGCN was selected to model relationships between customers and purchased products.

  2. Versioned data pipeline

    DVC records dataset and preprocessing versions alongside code changes.

  3. Tracked model experiments

    MLflow captures model parameters and ranking metrics for comparison.

  4. Small API serving surface

    FastAPI exposes recommendation requests without coupling clients to the training stack.

Evidence

What can be stated

Current status: Prototype. Unverified outcomes are labelled instead of presented as proof.

  • Model quality, dataset size, and latency figures remain unverified portfolio claims.

    Needs verification

Technology stack

  • PyTorch
  • FastAPI
  • DVC
  • MLflow

Next project

Continue through the work archive.

Upasana Centre

Start a project

Let's build something great

Drop your details below. We'll have it open in Gmail, ready to send — no copy-pasting.