BusyNow
Founder of BusyNow, helping people see what’s busy nearby before heading out.
It helps reduce the guesswork of choosing a place, while clearly showing when crowd information is recent, historical, or unavailable.
About my work on BusyNowls projects/
projects
Applied AI and predictive machine learning, from the problem to the implementation. Explore NFL forecasting, conversational agents, evaluation, and healthcare decision support.
// predictive machine learning
Fourth Down: NFL Predictive ML
An end-to-end, production-inspired NFL predictive ML application. Real current-season data, explainable frozen-model forecasts, historical evaluation, and an auditable forecast-to-outcome record.
Applied AI engineering · Frozen-model inference · Controlled improvement
// conversational AI
Sally
A healthcare voice agent designed around three workloads: a responsive conversation loop, background analysis, and an evidence-checked visit brief.
I lead its architecture and development, including shared browser and phone logic, tool-result validation, and scenario tests. The essay covers how it keeps conversations fast while making sure it never confirms an action that didn’t happen.
Live production · Patient outreach calls at pilot scale
Read the Sally essay 7 min// agent orchestration
Symphony
Engineering workflows that connect a bounded task to an agent-produced change, reproducible checks, and human review.
My work adapts OpenAI’s Symphony foundation to scoped context, specialized agents, isolated execution, and a separate publication boundary. The essay explains how agents get enough freedom to fix real code while authority, spend, and release stay under control.
In use · Built our products under human supervision
Read the Symphony essay 7 min// quality & evaluation
AI Evaluation
An evaluation system for healthcare chat and summaries that makes changes in model behavior visible, reproducible, and actionable.
I work across reference datasets, deterministic rules, independent model judging, repeated sampling, and paired scenarios. The essay explains how the Python and Bedrock harness separates evaluator correctness from model quality. It also covers how we investigate a failed result.
Live production · Gates real releases
Read the AI Evaluation essay 7 min// decision support
Referral Intelligence
A specialist assessment service that returns ranked candidates together with scoring contributions, evidence gaps, and reasons to withhold a recommendation.
I designed the boundary between assessment and workflow execution, with deterministic scoring and parity tests for the service extraction. The essay also covers my separate, in-development 30-day referral-completion model on synthetic data, and why ranking and prediction stay apart.
Live production · Ranking service
Read the Referral Intelligence essay 6 minsolution_design
I start with the customer workflow and the decisions an AI system needs to support. I map the context, tools, integrations, and outputs it requires, then define how to evaluate its behavior and where human judgment belongs.
I stay involved from solution design through implementation: shaping agent capabilities, building prototypes, reviewing code, and using evaluation results to refine the system.
More about my approach