Dave Liu — Portfolio
ML Research Engineer · Senior Data Scientist · Systems Builder
ML Research Engineer · Senior Data Scientist · Systems Builder
Skim the sections — or grab the resume PDF / resume page.
● Now — what I'm building and shipping this month
I knew I wanted to work in data before I knew what a data scientist was. At ten years old I noticed a pattern: civilizations chased gold and spices, then oil, and now—in the information age—data. It struck me that data is the modern gold, and that understanding the world at any real depth requires massive amounts of it. That idea has guided every career decision I've made since.
I believe the world is far more complex and nuanced than most people think—or care to see. Beneath every decision, every behavior, every interaction lies a web of signals that, if you look closely enough, tells a richer story than the surface ever could. To me, understanding those nuances isn't just intellectually satisfying; it's how we build a world that works better for everyone. Predicting the future starts with genuinely understanding the past and present, and I think the best version of our future begins by understanding people.
That philosophy has taken me across biotech, e-commerce, healthcare, recruiting, and fintech over the past decade. At Shipt I lead personalization for millions of grocery shoppers. Before that I helped build a cancer detection model at Freenome, designed a task-matching system that saved Change Healthcare $7M a year, and built talent-ranking models at Riviera Partners. On nights and weekends I run AutoTrader, a fully autonomous stock prediction system I built from scratch.
What ties all of that together is a belief that the hardest part of data science is rarely the model—it's getting the right data, defining the right metric, and making sure the thing actually ships. I spend as much time asking “should we be measuring this at all?” as I do writing training loops. It's an approach that's led to production systems at every company I've been part of, and one I don't plan on changing. In 2026 it became an independent research program—three self-published papers on multi-agent synthesis, model selection under distribution shift, and LLM-grounded feature extraction—all built on the same rule I use in production: the system that produces a number is never the one that grades it.
Lead data scientist on Shipt's Personalization team, architecting recommender and eligibility systems behind personalized shelves for millions of grocery shoppers. Mentored 4 data scientists, including 2 direct reports.
Core ML engineer at a genomics company developing a blood test for early-stage cancer detection. Worked at the intersection of infrastructure and research, building the systems that scientists depend on daily.
Worked on ML systems for health insurance claims processing—a domain where model accuracy translates directly into operational cost savings.
Built ML models for an executive recruiting firm, working across the full pipeline from data collection to model serving.
Two concurrent research positions exploring ML applications in energy and neuroscience.
Where the foundation was built.
"Dave is one of the most thorough and thoughtful engineers I've worked with. He doesn't just build models—he builds the systems around them that make sure they actually work in production."
Former Colleague
Freenome"What sets Dave apart is his willingness to ask the hard questions—about metrics, about assumptions, about whether we're solving the right problem. That intellectual honesty makes everyone around him better."
Former Manager
Shipt"Dave taught me git, and somehow made it make sense. He has a rare ability to explain complex technical concepts in a way that doesn't make you feel stupid for not already knowing them."
Research Scientist
FreenomeA research system (reference implementation “Metis”) for a recurring class of high-stakes problems: compressing a qualitative record—a court opinion, a clinical note, a field dispatch—into a calibrated empirical verdict. Large language models read such text fluently but fail here in two documented ways: they fabricate, and they degrade when allowed to judge their own improvements. This architecture confronts both. The LLM is confined to two roles—proposing a feature schema and performing verbatim-quote-grounded extraction—while a small, calibrated classical model makes the actual prediction; any extracted value without a verbatim evidence span is dropped and counted against a measured hallucination rate. The system self-improves only when a domain expert and a frozen gold set the loop never fits to both endorse a change, adjudicated by a permutation test with a hard subgroup-fairness veto.
Evaluated not by chasing a benchmark but as a known-answer test: across four real legal corpora spanning six decades of empirical judicial-politics findings, it independently re-derives the established signals (judge party predicts decision direction; repeat-player status predicts settlement) and the established nulls (judge attributes don't predict settlement; panel composition washes out). It exhibits the honesty its design targets—vetoing one of its own expert-approved escalations on held-out evidence, and on real European Court of Human Rights texts extracting ~1,200 features at a measured-zero hallucination rate yet landing at chance, the correct answer to a known leakage trap where prior text-only work reported ~79%. Every headline number is independently re-derived from scratch by an automated verifier (8/8 reproduce with zero drift).
READ THE WRITE-UP → READ THE PAPER → EXTRACTION-AUDIT TOOL →A multi-agent LLM framework where N expert agents—each with a unique professional persona, analytical framework, and domain-specific decision criteria—analyze queries in parallel, then a weighted synthesis step produces structured decision documents with confidence scores, dissent preservation, and risk matrices. Designed for high-stakes domains where single-model outputs are too overconfident and opaque.
Empirically evaluated across 750+ benchmark runs, 6 domains (medical, legal, business, policy, software engineering, general knowledge), and 5 models spanning 3B to frontier-class. Key findings: weighted synthesis outperforms single-best by 29–58% (p<0.0001, d=2.16), the optimal aggregation method is domain-dependent (synthesis excels in business at 100% accuracy, but single-best leads in legal at 75%), and synthesis amplifies model quality non-linearly with mid-tier models benefiting most (2.5–3x amplification). Published as an independent research paper with 47 citations across 13 related work categories.
TRY META COUNCIL → READ THE PAPER →What started as a weekend experiment to see if I could beat a simple moving-average crossover turned into a fully autonomous system that collects market data for 600+ tickers nightly, engineers 400+ features from eight distinct sources, trains 1,800+ dual models (one classifier for direction, one regressor for magnitude), and delivers confidence-ranked predictions to subscribers every morning before the market opens.
I built every piece myself: the data ingestion pipelines, a custom feature store spanning technical indicators to social sentiment, a dual-model training framework with walk-forward validation and Optuna hyperparameter optimization, FAISS-powered similarity search, a tiered email subscription system with Stripe billing, and the multi-cloud infrastructure (GCP + Azure) that orchestrates everything—all running autonomously for about $235/month.
EXPLORE THE SYSTEM →A pre-registered psychometric instrument that measures the gap between two channels of ethical preference: the values someone states they hold, and the values revealed by their everyday choices over weeks. Standard ethics measures capture one channel or the other; this is built to compare them within-person, longitudinally, without ever telling the user a choice was right or wrong.
The design spans four ethical domains and three scenario types (quick-fire timed choices, branching narratives, and cost-of-virtue stake ladders), scored with Bradley-Terry pairwise models and bootstrap confidence intervals across nine pre-registered hypotheses. It carries a growing family of novel measurement claims: narrative immersion as a debiasing mechanism (H8), self-prediction calibration (H9: do you know your own gap?), and design-stage work on cross-situational consistency, the moral-circle radius, the self–other double standard, an informant “moral 360,” a real-stakes validity probe, and a moral-language channel—each treating a different facet of self-knowledge (or of whether the instrument captures real behavior at all) as something measurable. A working local-first runtime now ships the daily practice on-device (your data never leaves the browser); a 61-scenario corpus with five recurring-character arcs is authored; OSF pre-registration (filed before recruitment) is in progress.
READ THE WRITEUP → TRY THE 90-SECOND DEMO →