Abhishek Nalla

Built the video product widget surfaced on Amazon detail pages — a high-traffic, latency-critical feature serving 10M+ requests/day. Solved the latency challenge using lazy loading and delayed REST calls so above-the-fold content renders instantly. Ranking and recommendation signals determine which videos surface per customer, driving $600M+ in increased revenue.

Built gRPC-based clip-generation and serving workflows for Amazon's AI shopping assistant Rufus. Core focus was ranking and relevance — surfacing the most useful video moment for each product question using ranking models, vector retrieval, and multimodal signals. Achieved 7% CTR.

Migrated DynamoDB read-heavy access paths to Elasticsearch for a Tier-1 search service. Designed shard strategy, tuned indexing and query relevance, and built ranking signals into search — cutting latency by 85% and enabling more accurate, personalized recommendations at 10% MoM traffic growth.

End-to-end async event-driven system for driver license verification using SQS queues and AWS Step Functions — eliminated manual review bottlenecks at scale.

Built RNN/LSTM & attention-based models to produce high-quality abstractive summaries. Trained and evaluated on a large corpus at IIIT Hyderabad's NLP Lab.

Trained Transformer models to learn symbolic math rules — simplification, factoring, and multi-step equation transformation — through structured reasoning tasks.

Event-driven webhook infrastructure supporting async callbacks for third-party integrations — reduced synchronous polling load and improved reliability across client APIs.