Research

I’ve spent a lot of time in school. Here’s where some of it went.

A large part of my work focuses on turning research into tools that can actually be used in practice. I draw on my software development background to build around the methods I develop, with particular attention to clinical and operational settings where the value of a model depends not only on its performance, but on how well it fits into real workflows.

Clinical decision support

Live operational pilot

Autism assessment screening tool

Turns scanned referral packages into a structured first review, bringing screening signals, supporting evidence, and assessment history into one workflow.

  • Next.js
  • FastAPI
  • OCR
  • Ollama
Traceable evidence
Extracts concern-specific evidence with source-page references, alongside structured referral fields and prior assessment history.
Human-in-the-loop design
AI organizes evidence and flags uncertainty; clinicians interpret the findings and make screening decisions. Indeterminate checkboxes and borderline referrals are surfaced for closer review.

Optimization / operations

Capacity planning & scheduling

A planning engine that translates care pathways, waitlists, and clinician availability into feasible schedules across individual and group treatment.

  • Python
  • Gurobi
  • Mixed-integer optimization
Grounded in stakeholder collaboration
Months of system understanding, workflow design, and stakeholder collaboration shaped the model’s requirements, from treatment spacing and group eligibility to capacity limits and aging-out deadlines.
Beyond fixed treatment blocks
Tracks treatment occupancy week by week to support repeated courses, then translates grouped optimization results back into individual client schedules.

Interactive learning

Agentic-AI scheduling simulation

A simulation environment where learners build scheduling workflows and explore the trade-offs between waiting, utilization, and capacity.

  • React
  • Python
  • AI agents
  • Docker
Composable decision workflows
Connects agent and Python blocks in a dependency-ordered graph, with scheduling benchmarks to put the resulting policies in context.
Execution with boundaries
Runs Python blocks in containers with network access disabled and resource limits. Proposed schedules are checked for feasibility before they can be applied.