Hi, I'm Afshin Shahrestani.

• Led the zero-to-one build of the CV/ML function; managed two ML engineers and partnered with CTO/CEO, backend, and product to define requirements and ship production code.
• Designed, implemented, and optimized transformer-based detection/segmentation and grounding pipelines for large-scale object processing, meeting strict runtime/latency constraints on CPU/GPU targets.
• Owned the ML lifecycle end-to-end—data ingestion, preprocessing, training/evaluation, packaging, CI/CD, and deployment—on GCP (Vertex/Cloud Run) and AWS (SageMaker/ECS); migrated the codebase to Hydra configs and Makefile-driven automation.
• Built high-throughput video and dataset pipelines (extraction, balancing, filtering, versioning) with major runtime improvements for multi-hour workloads; created scalable pre/post-processing flows tailored to field data.
• Implemented geospatial components and SfM/triangulation workflows to convert detections into accurate, map-aligned outputs for customer reporting.
• Developed and maintained internal Python packages used across repos—versioned, documented, installable—and introduced shared utilities for data, models, and evaluation.
• Established engineering standards, code reviews, and a monitoring approach (error buckets, drift checks, metrics) to support ongoing model reliability.

Stack: Python, PyTorch, HuggingFace, OpenCV, Hydra, Docker, GCP (Vertex AI, Cloud Run, GCS), AWS (SageMaker, ECS, EC2, S3), MongoDB, GitHub CI/CD.

• Built computer-vision models for 2D & 3D object detection/segmentation using PyTorch with YOLO, Detectron2, and MMDetection.
• Thesis focused on data generation for 2D detection to balance low-resource, imbalanced datasets; designed pipelines for augmentation/synthesis and class-balance curation.
• Developed reproducible training/evaluation workflows (experiments, ablations, visualizations) and documented results for downstream reuse.
• Published two papers (TAC 2024, IEEE 2025).

Stack: Python, PyTorch, YOLO, Detectron2, MMDetection, Computer Vision, GANs.

• Built and deployed real-time predictive models for financial time series (LSTM/RNN) to a live data feed with 5-second end-to-end latency; supported multi-granularity inputs (1-minute to 4-hour).
• Implemented data ingestion/feature pipelines, experiment tracking, reproducible evaluation harnesses; collaborated with engineering/product and conducted code reviews to ship production services.
• Explored reinforcement learning as an R&D track and reported findings/limitations; insights fed the model improvement roadmap and monitoring criteria.

Stack: Python, TensorFlow, LSTM, RNN, Reinforcement Learning, Time Series Analysis.