Lucas Brooks

Sets Applied Research direction for tabular foundation models and LLM-generated synthetic data, carrying that work through to published, open-sourced, production-usable form. Second author on PersonaLedger, a persona-conditioned LLM generation system paired with a programmatic engine enforcing accounting correctness through iterative feedback, producing 30M transactions across 23,000 synthetic use

Moved privacy-preserving synthetic data from tabular generators to LLM-based generation, and built the evaluation gates that let it clear model risk review. Fine-tuned Llama 3 with LoRA/QLoRA and DPO for structured financial record generation and extraction, improving exact-match extraction by ~15 points over the prompt-only baseline.

Hired to bring federated learning and privacy expertise into Capital One's machine learning research program. Publicly credited contributor to Federated Model Aggregation (FMA), Capital One's open-sourced federated learning framework, and co-inventor on a patent for federated learning utilizing customer synthetic data models.

Joined at spin-off to build the federated learning engine behind a privacy-first health data network, proving models could train on hospital and genomic data that could never be pooled. Built the federated training and secure aggregation layer for Equideum's Data Integrity and Learning Network (DILN), enabling joint model training across partner health systems while records stayed inside their ori

Developed the lab's data-analysis pipeline for evaluating time-resolved spectral estimation methods on neuroelectrophysiological recordings, replacing per-researcher ad hoc comparisons with a single reproducible benchmark. Applied specparam and neurodsp to analyze periodic and aperiodic spectral dynamics in ECoG data.

R&D on private and collaborative machine learning systems, work that seeded the Equideum Health spin-off. Prototyped collaborative training with PySyft and PyTorch, benchmarking federated averaging against centralized baselines on partitioned health datasets and quantifying the accuracy gap.

Built a regional blockchain community, running educational and collaborative programming for developers, entrepreneurs, and business owners, an entry point into the distributed-systems work that followed at ConsenSys.

Developed a cyclic well operation method that cut water cut from 90% to under 10% while reducing energy use; managed scheduling and accounting for a $5MM Chevron EMC project using Viewpoint and Excel, with AutoCAD Civil 3D and Trimble GPS for volume analysis and as-builts.

Master's in Computational Science with a focus on Machine Learning.

Bachelor of Science in Petroleum & Natural Gas Engineering.