Experience

Tickr, Inc. · San Francisco

• Own technical direction across product, engineering, and AI. Grew the function from one person to nearly twenty applied scientists and AI/data engineers across 10+ concurrent initiatives.
• Lead research and technical direction for RiskWise, licensed by major insurers and Big Four firms. Built proprietary agents and models that cut costs more than 20×.
• Post-trained a 4B risk-attribution model, halving errors against GPT-5.2 at nearly 20× lower inference cost.
• Built agentic signal discovery using a multi-armed bandit over AI agents, feeding a discrete-time hazard model of multidistrict litigation consolidation at 0.90 held-out AUC. Point-in-time backtests for 2022, 2023, and 2024 each put 16 to 32% of the top 25 ranked risks into later MDLs against a sub-1% base rate, 8 to 24 months early.
• Designed a Thompson-sampling bandit over natural-language research briefs run as frontier-model search episodes, with reward computed in code rather than by the model.
• Lead source-specific modeling across LexisNexis, CourtListener/RECAP/PACER/JPML, SEC/DOJ/FDA, PubMed, and UCC-1 records under licensing and permitted-use constraints.
• Established adversarial LLM-as-judge, capability, regression, and human-gold evaluation suites for actuarial, legal, and compliance review.
• Set research strategy on Where Web Search Breaks, a 30-task benchmark showing 10.4-point traceability and 3.3-point rigor gains over frontier deep-research systems.
• Serve as RiskWise's technical face, demoing to Fortune 10 through Fortune 500 companies and defending the product with technical buyers.

Tickr, Inc. · San Francisco

• Fine-tuned Qwen 3 with QLoRA/PEFT on eight H100s and bge-large-en-v1.5 for retrieval, reaching 98.1% F1 across four hierarchical datasets against 86.1% for GPT-4o and 86.6% for Claude 3.5 under identical conditions.
• Trained and served DeBERTa-XXL across 500+ product classes and fine-tuned bge-large-en-v1.5 with triplet loss, lifting downstream F1 by more than 0.03 and anchoring a multi-million-dollar contract.
• Built Generative Predictor Search over millions of economic series with retrieval, reciprocal rank fusion, and LLM reranking, cutting out-of-sample MAPE 15.6% and running 13× faster than exhaustive search.
• Shipped chatCPG in 2023, an early ReAct agent that selected its own data sources to answer analyst questions, and built and evaluated its RAG architecture.
• Applied early reasoning LLMs to Medicare Advantage benefit classification, reducing compliance risk.
• Designed high-signal datasets, blinded experiments, and benchmark protocols for LLM quality, RAG reliability, classification, forecasting, and reasoning against human and competitor baselines.

Subpoena Solutions (acquired by ZwillGen) · Joinder (acquired by Brightflag) · SpeedLegal

• Built the first MVP at Subpoena Solutions of a system that reads subpoenas, search warrants, Section 2703(d) orders, preservation demands, and emergency requests, drafts grounded objections, and runs six LLM classifiers grounded in 50-state law. Led AI development through the acquisition.
• Advised Joinder on AI strategy through its acquisition by Brightflag, and SpeedLegal through product launch.
• Lead AI strategy for benchmark design, post-training, supervised fine-tuning, systematic error analysis, and domain-specific evaluation, cutting error more than 50% against frontier baselines on 20+ tasks.
• Translate ambiguous legal operations problems into structured ML tasks, labeled datasets, decision thresholds, and stakeholder-facing analyses for high-volume document automation.
• Mentored engineers building self-improving agentic loops that optimize instructions rather than model weights, for privacy-sensitive clients whose policies prohibit fine-tuning.

University of California, Santa Cruz

• First author on NEO, a conditional GAN for photometric super-resolution improving measurement accuracy by factors of 2 to 10. Featured in the NVIDIA blog, April 2026.
• Co-author on a paper in Communications Earth & Environment (Nature Portfolio) on hypothesis testing and theory development in Earth system science.
• Built an agentic scientific discovery system that formulates hypotheses, designs and executes statistical experiments, performs adversarial self-checks, and iteratively refines a theory under MCTS-guided experiment planning.
• Train and run analysis on NASA Pleiades and NERSC supercomputers under Slurm.

Tickr, Inc. · San Francisco

• Built and deployed transformer architectures, Bayesian time series models, causal inference systems, and scalable ML infrastructure on Kubernetes, MLFlow, Snowflake, AWS, and Docker.
• Engineered NLP pipelines for entity extraction, sentiment analysis, relevance scoring, and unsupervised clustering; built labeling and quality systems using Amazon GroundTruth.
• Spearheaded the shift toward AI-first products and grew the technical function as the company scaled.

Market.Space (acquired by Tickr) · San Francisco

• First data scientist. Built and deployed customer-facing NLP models and production ML APIs using spaCy, scikit-learn, Flask, Docker, and AWS.

Salesforce · San Francisco

• Built churn prediction models and delivered statistical analysis correlating heterogeneous data sources to revenue.

NASA Ames Advanced Studies Laboratory · Moffett Field, CA

• Trained recurrent neural networks to predict energy microgrid load and demand from sensor data using Theano.
• Integrated microgrid sensor data with a Raspberry Pi and pushed it to a MySQL database for inference.

University of California, Berkeley

University of California, Santa Cruz