David Liu

Lead AI engineer (concurrent with Shipt) at a healthcare startup building behavioral-health solutions with LLMs; I own the AI validation and safety tooling — clinical-SME review loops, agentic evaluation, orchestration, and experimentation.

• Building the clinical validation framework: scores agent decisions and escalates uncertain or high-stakes cases to clinical subject-matter experts, converting each SME ruling into reusable development signal — the same expert-gated pattern as my Adaptive Domain Intelligence research in legal AI (LLM-as-judge, with changes shipping only when a domain expert and a frozen gold set agree).
• Built an LLM-as-judge pipeline that turns the varied visual elements in clinical documents into consistent determinations.
• Standardized agentic outputs across the product and validate them in a PHI-compliant AI Sandbox, with cross-agent orchestration on AWS Bedrock under a zero-retention configuration.
• Built SMOTE-inspired synthetic test-case generation — controlled perturbations of real examples — to expand regression coverage of already-observed failure modes and error classes without fabricating novel behavior claims.
• Formalizing an experimentation process for prompt changes: configurable candidate boards and A/B tests across multiple prompt variants, integrated with Langfuse for prompt registration and auditable, reproducible results.

Lead data scientist on Shipt's Personalization team; architected recommender and eligibility systems behind personalized shelves for millions of grocery shoppers. Mentored 4 data scientists, including 2 direct reports.

• Built and launched Deals For You V2, replacing matrix factorization with two-tower retrieval and precision reranking; a CUPED-corrected A/B test drove +91% shelf engagement, +1.76% order incidence, and +1.41% orders per user.
• Audited pre-treatment imbalance and a variant activation gap before SVP planning, correcting the annualized estimate from +83K to ~50K incremental orders (95% CI: 20K–80K) and ~$3.7M directional GMV; preserved a real win while retiring an inflated topline.
• The validated result prompted leadership to re-plan H2 around Deals as a potential ~$6M contributor to Shipt's $20M incremental-GMV target, prioritizing scale-up, holdback measurement, and reuse of the eligibility platform.
• Architected a reusable real-time Eligibility platform: a store-scoped hard pre-filter before ANN with swappable content tables; reduced an initial 448M-row design by ~95% to ~20M store-specific rows plus a compact global set, with zero-downtime refreshes and p99 <750 ms serving.
• Across the broader personalization portfolio, delivered up to 23% Personalized GMV lift across A/B-tested shelves and a cold-start strategy with +47% engagement and +8% first-time orders.
• Created the Personalization Interaction Score (views → clicks → ATC → purchases) and an automated Shelf Attribution pipeline; challenged an irreproducible 5% attribution claim, established the defensible 1–4% range, and raised the organization's measurement standard.
• Partnered with Engineering to replace CSV-based recommendation delivery with Kafka pipelines across legacy recommenders; served as the team's sole data scientist for its first four months while maintaining 16 repositories.
• Designed a Retrieval-Augmented Generation system over Shipt's retailer catalogs and built the business case that secured investment in an agentic-AI framework.

Core ML engineer at a genomics company developing a blood test for early-stage cancer detection. Worked at the intersection of infrastructure and research.

• Key contributor to Freenome's core multiomics cancer-detection model that predicts cancer stage (1–4) from blood-draw data; built data abstractions for petabyte-scale genomic datasets that unblocked cross-analyte feature development and accelerated training/evaluation cycles.
• Built a model-comparison system tracking research-vs-production model performance side-by-side — required for FDA audit compliance, gave the team confidence that production models stayed aligned with research intent.
• Designed and built large portions of Freenome's distributed ML training and serving platform used daily by 30+ scientists; scaled CPU-bound training across O(100) machines, supported leave-one-out / K-fold evaluation, and built reproducible model-artifact storage.
• Led adoption of PyTorch, MLFlow, and RayTune across the ML team — replaced legacy tooling to unlock GPU acceleration, experiment tracking, and hyperparameter tuning at scale.
• Built a cloud-cost monitoring system surfacing the biggest GCP storage and compute expenses; visibility alone drove optimizations saving the company over $10M annually — one of the highest-ROI projects I've worked on.
• Recognized with a Servant Leadership Award, elected by managers and peers across the engineering organization.

ML systems for health-insurance claims processing — accuracy translates directly into operational cost savings.

• Designed a ranking model matching human workers to claims tasks based on skill, history, and complexity; $7M annual value by reducing manual task assignment and the need for additional hires.
• Built a classification model partitioning sensitive patient documents (image + text) to route claims to the correct processing workflow.
• Developed internal AWS tooling and production API infrastructure for the ML team's deployment pipeline.
• Led a cross-functional tiger team prototyping a conversational chatbot (Rasa + HuggingFace NLP) for internal claims-inquiry workflows.

ML models for an executive-recruiting firm; full pipeline from data collection to model serving.

• Developed a model suite: a job-departure-likelihood classifier, a regression model predicting team sizes from resume features, and a candidate-ranking model using a custom NDCG listwise loss.
• Built an end-to-end framework for rapid model prototyping, training, evaluation, and serving — enabled the team to iterate on new models without re-engineering infrastructure each time.
• Wrote scrapers harvesting structured candidate data from public sites and APIs.

• California Institute for Energy and Environment (CIEE) — built a recurrent neural network for predicting building energy usage, exploring how temporal consumption patterns can inform smarter grid management.
• Bengson Research Lab, Sonoma State — applied ML to EEG data to predict individualized occipital lobe activation patterns; demonstrated early feasibility for brain-computer interface applications.

• Data Science Intern, Castlight Health (2017) — entity matching and deduplication pipeline using gradient-boosted classifiers with hard-negative mining; 85–95% precision/recall across hospital, facility, and practitioner entities.
• Data Science Contractor, Riviera Partners (2016) — team-size prediction model from public data; Python wrapper for survival-model time-series analysis; Flask model-serving infrastructure.
• URAP, Berkeley Institute of Data Science (2016) — mapped UC Berkeley course progression across majors via class-taxonomy organization and deduplication.
• Data Science Intern, Doximity (2015) — gradient-boosted classifier for malformed scraped articles; reverse geocoding + fuzzy string matching to link doctors in news articles to facility profiles.