Models designed to be validated, not just demoed.
We do applied research on recommendation and prediction systems, hybrid architectures, honest evaluation, and reporting your team can actually defend.
Where we spend our research time.
Graph learning
GNNs over real user–item graphs to capture relationships that flat models miss.
Sequence modelling
Transformers over behavioural event histories to model how interests evolve over time.
Recommendation & ranking
Top-K ranking, cold-start handling and cross-platform user-interest prediction.
Hybrid architectures
Fusing graph, sequence and tabular signals into a single model, the GNN + Transformer line.
Evaluation & ablations
Strong baselines and ablation studies so you know exactly what earns the accuracy.
Technical reporting
Clear write-ups: methodology, metrics, limitations, ready for stakeholders or review.
Research that survives contact with reality.
A model that wins on a slide but not on a baseline isn't a result. Every claim we make is backed by a comparison you can reproduce.
Define the model question
Turn a vague data problem into a measurable prediction task with a clear metric.
Build & compare
Implement the hybrid design alongside 7 baselines and 3 model variants; ablate what matters.
Report & hand off
Deliver validated findings, code and a written report you can act on.
Have a prediction problem worth researching?
Tell us about the data and the decision it should drive, we'll define the model question.