Shreyansh Agarwal

Led R&D for AI-assisted digitization workflows, integrating Segment Anything Model (SAM) into a geospatial platform for prompt-based, real-time segmentation of drone imagery.

Designed a high-concurrency RTI server using parallel, async WebSockets for low-latency, hover-based prediction interactions. Smooth like butter.

Optimized model pipelines for 2x performance improvement, enabling use of lower-cost GPU instances and significantly reducing infrastructure overhead.

Reduced end-to-end inference to sub-100ms by re-engineering SAM2 state management and freezing the image encoder. Speed is a feature.

Mentored a cross-functional team on production ML standards, leading the transition from research PoC to horizontally scalable cloud infrastructure.

Architected and delivered 10+ production ML workflows, consolidating fragmented models into a unified pipeline. 75% faster model onboarding.

Fine-tuned a ViT-based multimodal architecture using LoRA adapters, extending RGB vision pipeline with DEM modality via mid-fusion architecture.

Improved inference throughput by 33% by strategically merging LoRA layer weights back into base model. Fewer FLOPs, same quality.

Led hyperparameter optimization for complex unsupervised ML algorithms, achieving 20% direct improvement in model accuracy.

Implemented feature-specific post-processing algorithm: immediate accuracy ↑. Visual consistency for high-stakes client demos. 💼

Designed a cross-platform, Docker-based inference ecosystem with memory-safe tiling and batching for multi-gigabyte GeoTIFF processing.

Fine-tuned a YOLO-based object detection model for domain-specific targets, achieving 93% detection accuracy through dataset curation.