Skip to portfolio content
SYcontact

AI engineer · medical foundation-model researcher

Shuolin
(Leo) Yin

Toronto, Canada 43.6532° N · 79.3832° W

I build intelligent systems from research question to dependable product—across foundation models, multimodal AI, and autonomous agents.

Experience
01

Experience

Research depth.
Product velocity.

Sep 2026 – Present

Sanofi — AI Platform Team

Toronto, ON
  • Kubernetes
  • Amazon SageMaker
  • SageMaker Unified Studio
  • MLflow
  • AWS
  • CI/CD
  • Engineer the core of Sanofi's enterprise AI platform on the AI Platform team, building OneAI, SageMaker Unified Studio (SMUS), and the AI Foundry: the shared infrastructure on which data-science and ML teams across the global organization train, evaluate, and ship models.
  • Design and operate multi-tenant Kubernetes and Amazon SageMaker compute for distributed training and low-latency model serving, owning GPU scheduling, autoscaling, workload isolation, and cost governance.
  • Build the MLOps backbone on MLflow (experiment tracking, model registry, lineage, and promotion gates) so every model moves from notebook to production through a governed, reproducible, audit-ready path fit for a regulated pharma environment.
  • Ship golden-path templates, CI/CD pipelines, and self-serve environments that take teams from idea to a deployed endpoint without touching raw infrastructure, standardizing patterns across projects.
  • Harden the platform end to end with observability, reliability, and security controls, partnering with data-science, cloud, and security stakeholders to keep AI workloads dependable at enterprise scale.

May 2025 – Present

Vector Institute & University Health Network — WangLab

Toronto, ON
  • PyTorch DDP
  • DINOv3
  • H100 / SLURM
  • Qwen3-VL
  • MedGemma
  • Weights & Biases
  • Orchestrate large-scale experiments on HPC clusters with SLURM; posters accepted at IGCS 2026 (co-first author) and WCLC 2026 (second author), and contribute to manuscripts under review at Medical Image Analysis and Nature Communications.
  • Pretrain and finetune DINOv3 on 300M medical images using H100 GPUs with PyTorch DDP, achieving 10–20% improvements over SOTA.
  • Lead the MICCAI FLARE 2025 technical committee in designing evaluation metrics; release 3 fine-tuned Qwen3/2.5-VL and MedGemma models.
  • Present a self-supervised pathology foundation model at the NeurIPS 2025 SLCPFM Competition, outperforming SOTA across 5 downstream tasks and 3 cancer types.
  • Publish a medical vision-language models paper in the MICCAI Educational Challenge (top 5), open-sourced as an end-to-end training cookbook (200+ GitHub stars).
  • Collaborate cross-functionally with clinicians, translating medical domain knowledge into ML architectures tracked with Weights & Biases.

Jun 2026 – Sep 2026

Tsinghua University & Li Auto

Beijing, China
  • LangGraph
  • GRPO
  • Flask
  • KMeans
  • text-embedding-v4
  • Built an end-to-end evaluation platform for in-vehicle multi-turn dialogue, integrating conversation collection, generation, clustering, and evaluation in Flask over a simulated world of 23 cockpit subsystems.
  • Built a LangGraph agent that auto-generates multi-turn dialogues, optimizing human-likeness with GRPO reinforcement learning across 12 skill types spanning 4 domains and 3 modes.
  • Designed a dual evaluation framework quantifying tool-call accuracy (field/value-level F1 across 4 memory modes) and dialogue quality over 6 dimensions, packaged as a pip-installable evaluation library.
  • Built a user-persona Analysis Lab: embedded conversations with text-embedding-v4, clustered them with KMeans and LLM map-reduce, and visualized the results.
  • Lead-authored a survey on evaluating LLM-based conversational agents in the intelligent cockpit (submitted to Artificial Intelligence Review), built on a PRISMA-style 16-stage evidence pipeline with 260+ templated extraction records.

Jun 2026 – Sep 2026

Peking University

Beijing, China (Hybrid)
  • PyTorch
  • lifelines
  • Whole-Slide Pathology
  • Survival Analysis
  • Confounder Control
  • Led an H&E whole-slide pathology prognosis project centred on confounder control, reproducibility, and verifiable reporting rather than raw predictive performance.
  • Audited site leakage across the TCGA GI cohort (1,188 cases / 77 centres) with lifelines, exposing strong entanglement between centre and outcome (log-rank p < 10⁻³⁰).
  • Designed a leave-one-centre-out (LOCO) validation protocol, using the C-index gap between random splits and LOCO to quantify how much site confounding inflates apparent prognostic performance.
  • Built a WSI feature-extraction pipeline and tested whether centre identity can be predicted back from pathology features, localizing and blocking sources of site leakage.
  • Onboarded Beijing Cancer Hospital anti-HER2 and immunotherapy cohorts (with OS/PFS and centre labels) to support site-isolated external validation and H&E-based prediction of immunotherapy response.

Mar 2025 – Jan 2026

Medical3.ai

Toronto, ON
  • LangGraph
  • MCP
  • Citation-first RAG
  • SLM Distillation
  • MLOps
  • Built a multi-agent system for outbound sales ops covering discovery, evidence retrieval, outreach generation, and CRM follow-up with LangGraph and MCP.
  • Implemented citation-first RAG over FDA, CE/MDR references, ARTG, and internal documents, returning sourced excerpts for agent review and audit.
  • Distilled a task-specific SLM from Qwen3-32B, cutting inference cost by 7× while matching quality within 2% on automated + human evals (n=1,500).
  • Shipped a CRM automation agent with an MCP layer coordinating workflow across discovery, compliance, and outreach.
  • Supported pilots across 9 countries by shipping enrichment and retrieval features, reducing research time by 70% and lifting sales-team throughput.
  • Built an eval test set and guardrails; tracked retrieval hit and citation coverage, cutting unsupported claims by 30% via logging and error analysis.

May 2025 – Aug 2025

Lumi AI

San Francisco Bay Area (Hybrid)
  • OpenAI Agents SDK
  • LangChain
  • pgvector
  • Supabase
  • TensorFlow.js
  • iOS / React Native
  • Built HumanOS, a closed-loop autonomous agent that perceives, plans, intervenes, verifies, and learns across HealthKit, camera, and behavioral telemetry.
  • Implemented a tiered long-term memory system using Multi-Query RAG with MRR fusion over pgvector (1536-d) to drive persistent personalization.
  • Built conversational agent flows with OpenAI Agents SDK and LangChain; reached an average post-session user satisfaction of 9/10 across 500 beta users.
  • Shipped agent-driven interventions spanning VoIP, iOS FamilyControls, pledge gates, and multimodal verification, reducing average task delay by 40%.
  • Deployed behavioural analytics with TensorFlow.js and Supabase to track engagement metrics, enabling proactive, data-driven intervention.
  • Productionized infra on Supabase: PostgreSQL + pgvector with RLS, Deno edge functions, pg_cron jobs, and JS bridge handlers across native and React.

Feb 2024 – Sep 2024

Chinese Academy of Sciences — Institute of Automation

Beijing, China
  • Medical MLLMs
  • PyTorch
  • MLflow
  • GCP Compute Engine
  • Docker
  • Synthesized medical MLLM literature through systematic review and experimental methodology toward a TPAMI journal submission.
  • Constructed an end-to-end data pipeline for a multimodal evaluation benchmark with automated quality checks, reducing process time by 75%.
  • Developed a PyTorch multimodal evaluation framework assessing VLM accuracy, robustness, and cross-modal consistency, tracked with MLflow.
  • Coordinated interdisciplinary collaboration between medical experts and researchers, designing annotation protocols and improving accuracy by 40%.
  • Automated medical image labelling workflows in Python with NumPy and Pandas, systematically reducing validation effort.
  • Deployed evaluation infrastructure on GCP Compute Engine with Docker, enabling reproducibility across teams.

Jun 2022 – Sep 2024

VolunTrack Org.

Toronto, ON
  • TensorFlow
  • TensorFlow.js
  • Firebase / Firestore
  • React Native
  • React.js
  • CI/CD
  • Engineered a two-tower recommendation system in TensorFlow for volunteer matching, improving match acceptance by 85% on historical placements.
  • Built a real-time analytics dashboard with Firebase and TensorFlow.js, deploying engagement prediction models for churn analysis.
  • Developed a full-stack platform with React Native and React.js using a CI/CD pipeline for consistent, robust application builds.
  • Scaled the non-profit to 50+ members, 100+ partner organizations, and 500+ monthly active users through a four-tier management system.
02

Research

Clinical signal,
measured carefully.

01

First Author · 2025

From Concept to Code: A General Framework for Building a Medical Vision-Language Model

MICCAI Educational ChallengeTop 5
02

First Author · 2026

Autonomous Discovery of Multiple Instance Learning Architectures by Language Model Agents

Nature Biomedical EngineeringIn Preparation
03

First Author · 2026

Evaluating LLM-Based Conversational Agents in the Intelligent Cockpit: A Survey

Artificial Intelligence ReviewUnder Review
04

Co-First Author · 2026

Adapting Pathology Foundation Models for Prediction of Homologous Recombination Deficiency Status in High-Grade Serous Carcinomas Using Whole-Slide Images

IGCS 2026Accepted (Poster)
05

Second Author · 2026

LEADER: Lung Cancer EGFR, ALK, and ROS1 Detection Using Digital Pathology and Deep Learning

WCLC 2026Accepted (Poster)
06

Co-author · 2026

Self-Configuring Multi-Task Learning for Joint 3D Biomedical Image Segmentation and Classification

Medical Image AnalysisUnder Review
07

Co-author · 2026

Benchmarking and Adapting On-Device Large Language Models for Clinical Decision Support

Nature CommunicationsUnder Review
08

Presenter · 2025

Pathology-Native Self-Supervision at Scale: DINOv3 for Cancer Foundation Models

NeurIPS SLCPFM CompetitionPresented
09

Organizer · 2025

FLARE — Fast, Low-resource, Accurate, Robust, and Effectual Medical Image Analysis Challenge

MICCAIOrganizer
03

Writing

Notes from
the work.

01

August 14, 2026 · 14 min read

Building a room that behaves like an interface

A practical Blender-to-Three.js workflow for turning a composed 3D scene into a precise, expressive, and genuinely interactive website.

  • Blender
  • Three.js
  • Interaction design
02

August 12, 2026 · 8 min read

The half-life of an agent skill

The skills that patch today’s model weaknesses decay. The ones that carry context no model could know become durable infrastructure.

  • Agent skills
  • LLMs
  • Context engineering
04

Selected work

Systems built for
the real world.

01

Lead Engineer · Mar 2025

Echo – AI-Powered Sustainable Fashion Marketplace

Won 1st place in Canada and 2nd globally at BCG & Global Spark Hack the Globe, building an AI marketplace inside a 48-hour window.

  • React Native
  • YOLOv11
  • RAG / LangChain
  • GCP Cloud Run
02

AI & Full Stack Engineer · Apr 2024 – Jun 2024

EZ-Career – Autonomous AI Job Application Agent

Designed a multi-agent system on the OpenAI Agents SDK using an Agent-as-Tool architecture to orchestrate end-to-end job-application automation.

  • OpenAI Agents SDK
  • MCP
  • RAG
  • Supabase
03

Project Lead / Co-Founder · Jun 2023 – May 2024

YiXing – AI-Driven Personalized Travel Planner

Led development of an AI travel planner using fine-tuned GPT models with LangChain-powered RAG for personalized itineraries.

  • React Native
  • LangChain
  • AWS Lambda
  • GPT Fine-tuning
04

Project Lead · Sep 2024

ReassurED – LLM-Powered Emergency Care Navigator

Led a 24-hour hackathon build of a cross-platform emergency healthcare app using React Native (Expo) with real-time Firestore hospital data.

  • React Native
  • FastAPI
  • Few-shot Learning
  • Deepseek-v3