Raphael
Del Popolo
Software engineer — machine learning & generative AI
- Lead Software Engineer, Computer Vision at UTR Sports
- Previously LLM evaluation at Meta (via Wipro)
- Interim Secret clearance
- Remote
I build evaluation systems, data pipelines and on-device models for LLMs and computer vision. At Meta I built the LLM classifiers and judges behind an internal coding assistant and defined the failure-mode taxonomy its errors were counted against. At UTR Sports I am the sole engineer on a vision pipeline that turns a phone video of a tennis player into measured biomechanics, with 61,000 labelled pose frames now auto-labelled by a vision-language model and every model running on-device.
Selected work
Four projects that show the range: production vision, synthetic data, research, simulation.
The shelf
Every project as an object. Hover to read the label; click to open it.
Index
All projects, ranked by size and time invested.
Experience
UTR Sports
2026 — present · Remote
- Sole engineer building GenAI-powered coaching on an end-to-end vision pipeline: handheld phone video, outdoors, any camera placement, turned into measured biomechanics across three macOS apps, an iOS app and a watchOS companion.
- Natural-language and voice command layer on Apple Foundation Models on-device tool calling, with a deterministic parser behind the same tool interface.
- Qwen3-VL → SAM 3 auto-labeller that beat the hand labels on recall; owned the data preparation behind 61,000 labelled pose frames.
- Every model exported PyTorch → Core ML / ONNX on the Neural Engine; coaching-cue coverage raised from 30% to 81% under held-out, leakage-checked evaluation.
Meta, via Wipro
2025 — 2026 · Remote
- Evaluated Meta's internal LLM coding assistant across several product surfaces against a corpus in the low thousands.
- Improved its measured accuracy by 80% through controlled prompt and context experiments against a fixed test set; defined the failure-mode taxonomy used to classify its errors.
- Designed classification systems for thousands of LLM outputs and the SQL dashboards over them, cutting analysis time 90%.
- Escalation path for front-line reviewers: over 1,000 cases triaged, with the case-handling guides the team worked from.
Operation Care and Comfort
2023 — 2024 · Remote
- Built the platform matching donated goods to qualifying families, with eligibility from household profile, geolocation and reward history.
- Replaced hand-matching with eligibility rules and an allocation run that releases hundreds of event tickets at once.
- Generative AI
- LLM classification, grading and evaluation harnesses · retrieval and embedding pipelines · LoRA fine-tuning · vision-language models (Qwen3-VL) · on-device tool calling
- Data & evaluation
- Dataset design, versioning and review tooling · held-out bakeoffs against a measured noise floor · leakage audits · SQL dashboards over large evaluation runs
- Computer vision
- Detection, segmentation (SAM 2/3), keypoints, pose and tracking · auto-labelling · PyTorch → Core ML / ONNX on the Neural Engine, FP16, latency-profiled
- Stack
- Python · PyTorch · Swift 6 · TypeScript · SQL · OpenCV · Core ML · ONNX Runtime · SQLite · Git · Unix
- Education
- A.A.S. Artificial Intelligence & Machine Learning, and A.A.S. Computer Science — Chandler-Gilbert Community College · B.A. Acting & Film — University of Central Lancashire
- Also
- U.S. National Guard, Infantry · FEMA certified · German and Spanish fluent, French conversational · photography and videography