Speaking
Keynotes, paper presentations, workshop talks and programme-committee service. Each entry has its own page, with the slides where they are public.
2024
- Real-time Event Joining in Practice with Kafka and Flink
Paper presentation at the CIKM Online and Adaptive Recommender Systems workshop. Presented a production-grade approach to real-time event joining using Kafka and Flink, with war stories from running event streams that power a ~200M-user recommender system. Covered watermark handling, side-output patterns for late events, and cost/latency trade-offs between in-memory joins and materialized projections.
- Industry-track Program Committee
Served on the Industry-track program committee for the 33rd ACM International Conference on Information and Knowledge Management. Reviewed submissions on applied ranking, recommendation, and retrieval systems.
- Monitoring the Evolution of Behavioural Embeddings in Social Media Recommendation
SIGIR Symposium on IR in Practice paper presentation on how behavioural embeddings drift in a production short-video feed. Diagnosed silent quality regressions that offline retrieval metrics do not catch, and introduced embedding-space monitoring that tracks neighbourhood stability, drift against a reference snapshot, and coverage across cold-start cohorts.
2023
- Search, Recommenders and Co-pilots: Evolution of Evaluation Strategies in the Industry
Industrial-track keynote at the Forum for Information Retrieval Evaluation. Traced the evolution of evaluation methodology across search, recommendation, and emerging LLM-assisted co-pilot products — from offline relevance metrics to online multi-objective optimization and, increasingly, hybrid evaluation regimes for generative systems. Drew on production experience scaling recommenders at ShareChat from 1M to 200M users.
- Formulating Video Watch Success Signals for Short-Video Recommendation
Workshop paper presentation on principled formulation of the "success" signal for short-video consumption — distinguishing engagement from satisfaction, handling variable video length, and building a label that ranking models can learn against without overfitting to completion-rate artifacts.
- Building the Scalable Min-views Pipeline with HAMSA
Design walkthrough of HAMSA — vector-similarity-search for candidate generation in a high-throughput short-video feed. Covered the index layout, sharding, cold-start handling, and the migration that moved candidate generation from ANN-over-KVs to a purpose-built serving tier.
2022
- Production Recommender Systems at ShareChat
Community webinar walking through the architecture of ShareChat's recommendation stack — candidate generation, ranking, feature pipelines, and the ML framework investments that let a small team ship changes at high cadence.