Sibo Zhu朱思博
Hey, glad you made it here. 👋
I'm Sibo (/ˈsiːboʊ/). At Innoam Asset Management in Shanghai, I research systematic equity alpha strategies. Within six months of crossing over into quantitative finance from zero domain background, I took my first strategy all the way from initial research to live production deployment — work that spans signal discovery, model development, and the reproducible training, backtesting, and execution workflows needed to evaluate them. What I care about most isn't whether a model wins a single backtest, but whether an apparent improvement holds across periods and reproduces cleanly under real point-in-time data and operating constraints. My previous chapter: startup founder — co-founder & CTO of Cyber Partner AI, where our first AI companion, JiFeiFei, sold 10,000+ units in month one and passed 100,000 within half a year. Different battlefield now, same sport: building models that survive contact with reality.
Before that, I spent years in the trenches of AI and robotics. At MIT's HAN Lab, I worked with Prof. Song Han on making deep neural networks run fast and lean on edge hardware — techniques that later became standard industry practice. As Perception Lead of MIT Driverless, I led a 20+ engineer team, pushed perception accuracy (mAP) from 67% to 89%, and along the way pitched Google, Waymo, and Oracle into putting their logos — and $200K+ in sponsorships — on a student-built autonomous race car. Good times.
The part I enjoy most is dragging research out of the lab and into the real world. I've worked both ends of that spectrum — publishing at ICRA and IROS on one side, shipping hardware with real revenue on the other — and everything in between, from CUDA kernels to go-to-market execution.
Right now I'm happily heads-down on financial markets, but the door for good conversations never closes — technical deep-dives, startup war stories, research ideas, or ambitious builds. If you're working on something interesting, I'd genuinely love to hear about it.
Off the clock: snowboarding (AASI Level-II certified instructor — yes, I can legally teach you), photography, an overly engineered approach to meal prep, and poking at the boundary between the digital and the physical world.
The track record and portfolio below have the full story. Want to talk? Contact info is at the bottom — always happy to chat.
Track Record
Startups, labs, and everything in between
- Crossed over into quantitative finance from zero domain background and took a daily equity alpha strategy from initial R&D to live production deployment within six months, while continuing to drive signal blending and dynamic weighting enhancements in matched historical backtests with standalone execution packages
- Built reproducible rolling-training and strategy execution workflows, enforcing strict model versioning, point-in-time (T-1) data audits, and deterministic cold-start, state-recovery, and output-consistency verification
- Engineered an end-to-end automated factor research and governance pipeline, linking large-scale candidate generation, redundancy filtering, feature standardization, and paired model validation to evaluate cross-period stability and marginal portfolio contribution
- Developed novel features capturing multi-scale price dynamics and cross-sectional relationships, delivering standardized, model-ready factor libraries and evaluating their conditional incremental value against baseline factor pools
- Co-founded the company and led product & engineering: shipped our first AI hardware product to 10,000+ units in month one and 100,000+ within six months
- Architected the AI JiFeiFei companion device in partnership with leading children's content platform "Kaishu Story", squeezing 0.9s end-to-end latency onto an ESP32-S3
- Delivered an enterprise-grade 2-meter humanoid interactive robot (Agibot A2 + custom multimodal stack) for Volcano Engine's flagship exhibition
- Built and led cross-functional R&D across embedded firmware, multimodal LLM orchestration, and hardware mass production while securing strategic funding
- Wound down operations in mid-2025 and spent the transition period retooling for quantitative finance
- Researched event-based camera perception for high-speed autonomous driving under extreme lighting and motion conditions
- Designed sparse Vision Transformer architectures for asynchronous, microsecond-resolution event stream processing
- Co-developed SemAlign (IROS 2021), a self-supervised sensor calibration framework cutting camera-LiDAR calibration from 3+ hours to under a second (10,000x speedup)
- Built an efficient end-to-end LiDAR navigation system (ICRA 2021) deployed on a full-scale autonomous vehicle
- Collaborated with Prof. Song Han on edge neural network compression (PatchNet) adopted across industry
- Led a 20+ engineer perception team and co-pitched $200K+ in sponsorships from Google, Waymo, and Oracle
- Pushed perception mAP from 66.97% to 89.35% within a 200ms full-stack latency budget, enabling championship-winning track performance
- Open-sourced the perception and training infrastructure, adopted by 100+ Formula Student teams globally
Research Assistant
Brandeis University & Boston University
Massachusetts
- Developed a CNN-based motion blur detection system achieving 92% accuracy, open-sourced with 100+ GitHub stars
Education
Ph.D. Studies, Aerospace Science and Engineering (left to co-found Cyber Partner AI)
University of Toronto·2021 - 2023
M.S in Computer Science (AI/ML Specialization)
Brandeis University·2018 - 2020
B.A in Computer Science + Pure & Applied Mathematics (Double Major)
Boston University·2014 - 2018
Portfolio
Research, robots, and things that actually shipped
All of this is from my AI & robotics years — the optimization instincts, data pipelines, and ship-it discipline now go into markets.
AI JiFeiFei: 10K+ Units Sold in Month One
Our first product at Cyber Partner AI: an interactive AI companion for 'Kaishu Story', China's largest children's content platform. Sold 10,000+ units in the first month and passed 100,000 within six months. The technical core was squeezing 0.9-second end-to-end latency out of an ESP32-S3 via custom model compression and hardware-AI co-design, paired with long-term memory so it actually remembers the child it talks to.
Team: Sibo Zhu, Cyber Partner AI Team
April 2025
Bumblebee Interactive Robot: Enterprise Multimodal AI
A 2-meter humanoid robot for Volcano Engine's flagship exhibition hall, built end-to-end on the Agibot A2 platform with our custom conversational AI stack. Speaks two languages, stays in character, and survives the harshest test environment in robotics — the general public, 24/7. Vision, voice, and NLP stitched into one coherent persona, delivered on a non-negotiable exhibition deadline.
Team: Sibo Zhu, Cyber Partner AI Team
December 2024
SemAlign: Sensor Calibration in Under a Second
Co-first-authored calibration framework (IROS 2021), born from the chore everyone on the race team quietly hated: self-supervised semantic alignment — zero manual labels — that cut camera-LiDAR calibration from 3+ hours to under one second, a 10,000x speedup. A workflow the entire AV industry treated as a necessary evil became a non-event.
Authors: Zhijian Liu, Haotian Tang, Sibo Zhu, Song Han
October 2021
MIT Driverless: Autonomous Race Car & Open-Source Ecosystem
Led the 20+ engineer perception division for MIT's autonomous Formula Student race car while co-leading sponsorship pitches that secured $200K+ from Google, Waymo, and Oracle. Beyond pushing detection mAP from 66.97% to 89.35% on track, we packaged and open-sourced our dataset, labeling pipeline, and training infrastructure — now a foundation for 100+ student racing teams worldwide.
Authors: Sibo Zhu, MIT Driverless Team
August 2021
LiDAR End-to-End Navigation: Full-Scale Vehicle Deployment
End-to-end LiDAR navigation framework co-authored with Daniela Rus and Song Han (ICRA 2021). Sparse convolution optimization plus hardware-aware architecture design, demonstrated on a full-scale autonomous vehicle — not a simulator, the real thing. From raw point cloud to steering command in one pipeline.
Authors: Zhijian Liu, Alexander Amini, Sibo Zhu, Sertac Karaman, Song Han, Daniela Rus
May 2021
PatchNet: Ultra-Compact CNN for Edge Computing
Ultra-compact CNN co-authored with Huizi Mao (NVIDIA Edge AI Lead), Bill Dally (NVIDIA Chief Scientist), and Song Han (MIT): exploiting temporal redundancy in video to get down to 58M FLOPs — 5x fewer operations and 4.3x faster on edge devices at the same accuracy. Open-sourced, with its core design adopted across the industry.
Authors: Huizi Mao, Sibo Zhu, Song Han, William J Dally
March 2021
High-Speed Autonomous Racing: 200ms Full-Stack Vision
Co-authored IROS 2020 paper on practical computer vision for autonomous racing: a deeply optimized YOLOv3 detector, custom keypoint pose estimation, and sub-microsecond stereo camera synchronization, squeezed into a 200ms end-to-end latency budget at racing speeds.
Authors: Kieran Strobel, Sibo Zhu, Raphael Chang, Skanda Koppula
July 2020
Earlier Work
Faster LiDAR: Real-Time Point Cloud Processing Framework
November 2020
Attention-based transformer framework for predicting missing LiDAR frames in real-time autonomous driving. 40% throughput improvement with detection accuracy (mAP) unchanged.
PointPainting: Sensor Fusion Framework Implementation
August 2020
Implemented nuTonomy's camera-LiDAR fusion model PointPainting, then used it as the evaluation testbed for our 3D point-cloud prediction and sensor calibration research.
Grad-CAM: CNN Interpretability Framework
March 2019
Implemented Grad-CAM to visualize CNN decision-making, used for debugging the race car's object detection — in safety-critical settings, models need to be explainable.
Motion Blur Detection: CNN-Based Computer Vision
May 2018
CNN for motion blur detection in high-speed camera systems, 92% accuracy. Open-sourced it and picked up 100+ GitHub stars — my first taste of shipping something strangers actually use.
Get in Touch
The inbox is open — bring something interesting
"Knowledge only comes from empirical observation, everything else is just speculation." - Devon Eriksen