Honorable Mention Best Developer Tool
More than 550 students from universities across North America took part in MHacks 2024.
Zichen "Charlie" Zhang
About Me
I was the first full-time hire at Mira (formerly Halo), an AI smart glasses startup in San Francisco. I joined at pre-seed. Mira raised $6.6 million from General Catalyst, Naval Ravikant, Pillar VC, Soma Capital, Village Global, and angel investors.
At Mira, I worked on the voice and memory systems that turn everyday conversations into useful context for the glasses. That included real-time transcription, speaker recognition, backend infrastructure, and long-term agent memory.
Before Mira, I worked at Supercell’s AI Lab. Supercell is the company behind Brawl Stars, Clash Royale, and Clash of Clans. I built an LLM-powered NPC dialogue system whose characters responded to players with changing moods and branching storylines.
I studied Computer Science at the University of Michigan and was part of LSA Honors.
You can reach me at charliezhang325@gmail.com.
Projects
*
đ Best Extended Reality (AR & VR) đ. Lightspeed 2026 Game Changers
Mira is a pair of camera-free AI display glasses that turns conversations into searchable memory and provides real-time answers and translation. The 39g glasses have dual MEMS microphones, dual speakers, a private dual-waveguide display, and a smart ring for gesture control. As Mira's first full-time hire, I built speech-to-text, voice activity detection, and speaker diarization with Soniox and WebRTC; deployed voice-fingerprinting models with TorchServe and PyTorch Audio; and developed agent memory with mem0, semantic search, and graph retrieval. I also built FastAPI and WebSocket services on AWS ECS with autoscaling.Reward functions decide whether a reasoning model is pushed toward correct answers, shorter chains, or both. We trained Qwen2.5-3B with GRPO using normal, cosine, and dynamic rewards, asking whether length penalties that help larger models also work for small ones. In our runs, the normal reward steadily improved MATH500 and GSM8K accuracy. Cosine and dynamic rewards shortened the reasoning but made accuracy worse and less stable.
People learn about objects by both looking at and touching them. VTMo learns a shared representation of what objects look and feel like so autonomous agents can interact with the physical world using more than vision alone. The model uses shared self-attention plus modality-specific and cross-modal experts, combining the speed of a dual encoder with the cross-modal reasoning of a fusion encoder. Trained with InfoNCE on 3,600 image-touch pairs from Touch and Go, VTMo reached 57.27% Recall@1 on image-to-touch retrieval, compared with 15.11% for the frozen-attention baseline, while using fewer FLOPs than a CLIP-style dual encoder.
Tree-of-Thoughts can solve hard reasoning problems by exploring and backtracking across multiple paths, but its repeated prompts are expensive and poorly suited to small models. We compressed it into one structured prompt and distilled GPT-4o outputs into SmolLM-360M. On Game of 24, One-Step ToT reached 19% accuracy versus 7% for Chain-of-Thought, while fine-tuning on 144 examples raised SmolLM from 1% to 9%. Our full GPT-4o Multi-Step ToT run reached 82%, above the prior 74% GPT-4 result.
Learning Pushing Dynamics for Arbitrary 2D Rigid Bodies
Robots in homes and warehouses need to push irregular objects without a hand-built physics model for every shape. We trained dynamics models on simulated Franka Panda pushes so an MPPI controller could predict each object's motion and plan around obstacles. On L-, U-, and M-shaped objects, a deeper point-cloud-inspired network modeled pushes better than a shallow MLP and completed pushing and obstacle-avoidance tasks in fewer steps on average.
đ Fifth Place, Supercell AI Hackathon đ
Meowtive is a cozy-game AI system built with Unity, LangChain, and Google Gemini. Its cat agent chooses activities such as fishing or exploring based on its personality, mood, and surroundings. It can start conversations, ask the player for help when it gets lost, and carry out requests. A persistent memory tracks what the player enjoys and what the cat has learned from earlier mistakes.đ Honorable Mention, Best Developer Tool đ. MHacks 17 Hackathon, 2024
GenHint is a VS Code extension that teaches students how to work through a programming problem instead of returning a finished answer. It turns a comment into a code template with TODO steps, then explains each step and reviews the result. Our four-person team built it in 24 hours with TypeScript, the VS Code API, Groq, and Llama 3 70B, then published it on the Visual Studio Marketplace.Collage is an AI course discovery and planning tool for University of Michigan students. Students sign in with their school account, browse a Pinterest-style course catalog organized around their interests and career plans, ask an AI advisor for scheduling help, and share schedules with classmates.
MIA-Sort is a Python bioinformatics tool that extracts and sorts chromatin complexes from large Hi-C and Pore-C datasets. Researchers can use the output to analyze chromatin loops, stripes, jets, and hubs when studying loop extrusion.
Blogs
*
UMich EECS 482 Operating System Notes
General programming patterns for writing multi-threaded programs and implementing synchronization primitives with low-level hardware
A Deep Learning Tutorial Written for Everyone
A detailed introduction to the most common deep learning model architectures and theories, including Transformer, BERT, GAN, Meta Learning, Adversarial Learning, Reinforcement Learning, Network Compression, Anomaly Detection, etc.
Experience
Engineering
Stealth Startup
Privacy-focused voice agent with long-term IRL memory for display smart glasses
Venture Partner
Connecting startups and founders with funds and technical talent
Founding Software Engineer
Joined at pre-seed as the first full-time hire. Mira went on to raise $6.6 million from General Catalyst, Naval Ravikant, Pillar VC, Soma Capital, Village Global, and angel investors.
Built the real-time voice stack with Soniox, WebRTC, TorchServe, and PyTorch Audio, covering speech-to-text, voice activity detection, speaker diarization, and voice fingerprinting.
Developed the agent memory layer with mem0, semantic search, and graph retrieval. Built FastAPI microservices with WebSockets on AWS ECS and configured autoscaling.
AI Engineer Intern, AI Innovation Lab
Built an agent-based NPC dialogue system for the company behind Brawl Stars, Clash Royale, and Clash of Clans. The system generated dialogue trees at runtime and adapted each character’s mood and storyline to the player’s actions.
Developed the backend with FastAPI and the Google GenAI SDK, deployed it on AWS Lambda, and connected it to Unity C# game clients for real-time dialogue.
Co-Founder
AI personalized academic advising and scheduling at UMich
Research Intern, Minji Lab
Open-source tool for sorting massive chromosome datasets
Research Intern, Direct Brain Interface Lab
Automated user interaction for internal BCI survey tool
Awards
University Honors
Received in all six of my full-time terms at Michigan, from Fall 2022 through Winter 2025.
James B. Angell Scholar
Awarded for an A record across two or more consecutive terms. Received in 2024 and 2025.
William J. Branstrom Freshman Prize
Awarded to first-term students who rank in the top 5% of their class.
Fun Facts
- My Chinese name is åŧ į´Ģ厸. į´Ģ means purple, while 厸 originally referred to an emperor's residence.
- I grew up in Zhangjiagang, near Suzhou, China. I attended the International Department at Jiangsu Provincial Liangfeng Senior Middle School before transferring to Rochester Adams High School in Rochester Hills, Michigan.
- I've practiced Chinese calligraphy since elementary school. Some of my recent work is below and on Instagram.
Calligraphy
Photography
