Cut backend load 70% on a system handling 150,000+ daily transactions — swapped synchronous partner commission math for concurrent Python asyncio, chosen for true parallelism over caching.
Built a mass merchant onboarding API that collapsed a 3-day manual approval process into 30 minutes, scaling to 8,000+ merchants and feeding ~30% gross revenue growth.
Designed modular Django REST APIs for partner-merchant config, reused across teams and partners — serving 50,000+ partners at 98% uptime.
Shipped the company's first AI support chat on Amazon Bedrock, trained on internal policy, deployed across multiple AWS availability zones.
Cut deployment time from 2 days to 15 minutes with CI/CD pipelines using Docker and Jenkins, across 10 services.
Extended an interpretable transformer to 4-class Alzheimer's staging across 44,000 MRI scans, closing a gap present in over 80% of existing AD literature.
Used Triplet Attention to surface clinically relevant brain regions — explainability built for clinical decision support, not leaderboard score.
Trained on an NVIDIA A100, benchmarked against 4 CNN baselines — the twist: transformers need multimodal or volumetric input to match CNNs on 2D-only MRI tasks.
Real-time Spark Structured Streaming pipeline over live GeoJSON telemetry from Dublin's Moby e-bike fleet, using event-time windowing and watermarking to eliminate 14% duplicate GPS records.
Fault-tolerant aggregates written to Parquet on S3 and MongoDB, surfacing demand hotspots and station crowding up to 180–200% capacity.
Built with no trip-level data — every insight inferred from station-state changes alone, delivered through an interactive Streamlit dashboard.