Get in Touch

• Lead the data science effort on ALAI, a production LLM platform over financial advisor and client call transcripts, live to the full advisor field. Direct a small team on architecture, prompt design, and evaluation. Built on LangChain and LangGraph over AWS Bedrock.
• Built Memex, the in-house agent system the platform runs on. Confidential material cannot leave the firm’s platform, so everything runs and is stored inside it.
• Set the evaluation standard ALAI ships against: an LLM-as-judge harness as a fast pre-screen and never as ground truth, a rubric built with subject matter experts, and blinded comparative review of competing variants. One of the four modules reaches 93.4% field-level recall against expert-labeled transcripts.
• Designed ALAI Insights, a multi-agent research framework on the same stack. Each run moves from a stated question and hypothesis through investigation to a validated conclusion, with entity citations into a knowledge graph and context that continues across runs. Internal teams use it to find product gaps and the advisor behaviors that precede a client commitment.
• Built a two-stage financial planning assistant: an FT-Transformer for product propensity feeding a FiLM-conditioned transformer that predicts plan pages, trained on historical plans, so a partial plan can be completed from client features and the selections already made.

• Co-built a clothing recommender on Doc2Vec embeddings and gradient boosting. Customers kept 34% of the items it ranked highest, against about 17% for the rest.
• Replaced a snake-style warehouse picking route with an optimized one. Replaying the same orders, walking distance dropped over 24%.
• Built a model predicting which order items would be refilled from returns within 24 hours.
• Built Visual Sales, a dashboard ranking product imagery by best sellers.

• Worked cross-functionally to extract, transform, and load (ETL) customer data into Penta Technologies ERP database.

• Constructed predictive models using supervised machine learning algorithms to predict patient mortality risk.

• Led a remodel program comprising 348 plumbing supply stores across the US and expedited the timeline by 75%+.

• Designed and built an autonomous RL-based portfolio management system that ingests multi-source financial data (FRED, SEC, news, market), processes it through 15 specialized feature engineering modules, and outputs risk-managed portfolio allocations using PPO with a Transformer policy network.
• Implemented cross-asset attention mechanism, market regime detection, and reward shaping balancing return incentives against drawdown and turnover penalties.
• Integrated as a tool within an LLM-orchestrated agent: the model surfaces insights and allocation recommendations, discusses rationale with a human-in-the-loop, then executes approved actions.

• Built a knowledge graph–based task management system where a single logged action recursively propagates through a multi-parent taxonomy spanning 6 category hierarchies and 1,600+ nodes.
• Engineered adaptive frequency system using golden ratio–based adjustments: tasks completed early tighten in frequency, tasks completed late loosen — the system evolves to match actual behavior without explicit configuration.
• Trained prediction model achieving ~75% Hit@1 accuracy for next likely action, served as a tool to an LLM agent that interprets predictions, recommends scheduling changes, and executes adjustments after human-in-the-loop approval.

Shane Conner

Shane Conner

Shane Conner

Shane Conner

Shane Conner

Shane Conner

Shane Conner

Shane Conner

Shane Conner

Shane Conner

Patrick Barranger, Rohit Nair, Rob Mulla, Shane Conner

• Randomized online survey experiment (n = 495) demonstrating that merely assigning ownership of an airline seat amenity produces a 15–20x gap between willingness to accept and willingness to pay (p < 0.05), without possession or interaction.

Summa cum laude

Cum Laude · Swimming & Diving Team