LIA – Large Interview Assistant
A multi-agent AI interview coach delivering real-time feedback with language and computer vision analysis. I led full-stack development and integrated Google Gemini for adaptive question flows.
Hi, my name is
Josh Rauvola.
AI engineer & researcher building responsible, production-ready systems.
I design and deploy large language models, multimodal AI, and data products that help teams ship smarter, greener technology. Currently at U.S. Bank shaping human-centered AI in highly regulated environments.
Scroll
01. About
I’m an AI engineer with 4+ years of experience translating cutting-edge research into products that scale. I hold a Master’s in Applied Data Science from the University of Chicago and dual B.S. degrees in Computer Science and Economics from Cal Poly.
At U.S. Bank, I build machine learning solutions ranging from customer analytics and anomaly detection to NLP-powered knowledge tools—delivering over $10M in business value across finance and healthcare. I specialize in large language models, natural language processing, multimodal fusion, and time-series forecasting.
Highlights include leading the Fair Developer Score research (published at ASE 2025), winning the UChicago AI Hackathon with a graph neural network project, and developing LIA, a multi-agent AI interviewing coach that blends language and vision models.
Beyond work, you’ll find me exploring Green AI practices, contributing to open-source, and chasing PRs as a former track athlete (4:17 mile).
- Python
- PyTorch
- Transformers
- LangChain
- Computer Vision
- LLM Ops
- Docker
- SQL
- GNNs
02. Featured Projects
Fair Developer Score
A build-adjusted productivity metric that balances developer effort and impact, grounded in commit clustering and knowledge graph centrality. Co-authored and led experiments; currently published at ASE 2025.
ThoughtTrim
A chain-of-thought trimming technique that ranks and preserves anchor thoughts to cut 60–90% of tokens while maintaining answer accuracy. Built during Apart Research 2025 with Andrew Briand.
03. Journal
SyncSense: Music Supervisor Agent
Closed-loop scoring that binds emotion, music, and AI visuals—includes the full deck and presentation.
Watch and readThoughtTrim: Anchor-Driven RL Modification
How trimming low-influence steps in an LLM’s reasoning can recover efficiency, interpretability, and safer agentic behaviors.
Read postFair Developer Score: Build-Adjusted Measurement
Designing a productivity metric that celebrates meaningful engineering impact instead of noisy commit counts.
Read postLIA: Large Interview Assistant
AI-powered interview preparation with multi-agent LLMs delivering tailored questions and real-time coaching.
Read post