Skip to content

Details

Join our virtual meetup to hear talks from experts on cutting-edge topics across AI, ML, and computer vision.

Date, Time and Location

Jan 21, 2027
9:00 AM - 11:00 AM PST
Online. Register for the Zoom!

Verifier-Gated Generative Tutoring: When the LLM Writes Lessons but Never Decides Mastery

Adaptive tutors increasingly let LLMs both generate content and decide mastery — a risky coupling. dino.study separates those roles: models draft interactive lessons under verifier gates, while deterministic Python BKT + FSRS tracks mastery and review timing. I'll show the production architecture (Expo app + Bun API + Postgres prerequisite graphs), cost (~$0.04/lesson), and an honest LLM-simulated N=30/arm dry-run that is not a human peer-reviewed study.

The talk is code-forward with a live demo. Aimed at ML/CV practitioners curious about evaluation, mastery modeling, and shipping educational AI.

About the Speaker

Animesh Alang is a UCSC Computer Engineering junior building dino.study. He founded SlugAI, serves as President of CruzHacks, and was Cloud Track winner at NVIDIA Hack-a-Claw.

Building CAD for Machine Intelligence: The Future of AI-Driven Engineering

What does it take to build AI systems that can reason about geometry, design mechanical parts, and automate engineering workflows? In this talk, we introduce Quartz, a CAD system built for machine intelligence, and explore how high-performance geometry execution, reinforcement learning, and massively parallel environments enable AI agents to learn engineering through design and iteration.

We demonstrate how AI can create and modify complex assemblies while preserving editable CAD histories, and share our vision for the future of AI-driven engineering.

About the Speaker

Debao Huang is currently a Research Scientist at Spectral Labs, where he develops AI-native CAD systems and foundation models for automating engineering design. He obtained his Ph.D. in Electrical and Computer Engineering from The Ohio State University.

Does 70% Mean 70%? Building an Honest Tennis Win-Probability Engine

When a model says a tennis player has a 70% chance to win, does that player win 70% of the time? I built Tiebreak, a win-probability engine trained on about 900,000 ATP matches that layers surface-adjusted Elo, a point-by-point Markov chain and gradient boosting, updating live as a match unfolds.

Early versions looked better than they were because the same match leaked into both training and test data, and I show exactly how I caught it with a chronological, match-level split. Reliability diagrams then show where the model is calibrated and where it is overconfident.

Attendees leave with a practical checklist for building and honestly evaluating any ML model that outputs probabilities.

About the Speaker

Krrish Ghindani is a Data Science junior at the University of San Francisco on a full merit scholarship and leads USF's Data Science team at national competitions (1st place).

Can an Open-Weight Model Power a Physical AI Agent? Serving Gemma 4 on Google TPUs with Voxel51

In this session, we'll deploy Google's Gemma 4 on Cloud TPUs using vLLM-TorchTPU and connect it to the Voxel51 Agent. Through a live demo and inference benchmarks, we'll explore how open-weight models can support Physical AI workflows, examining latency, throughput, and the infrastructure tradeoffs behind building reliable AI agents.

About the Speaker

Adonai Vera is a Machine Learning Engineer and DevRel at Voxel51 with over 7 years of experience building computer vision and machine learning models using TensorFlow, Docker, and OpenCV.

Related topics

AI/ML
Artificial Intelligence
Computer Vision
Machine Learning
Robots

Sponsors

Voxel51

Voxel51

Administration, promotion, giveaways and charitable contributions.

Voxel51

Voxel51

Administration, promotion, giveaways, charitable contributions.

You may also like