███████╗██████╗ ██╗ ██████╗ ███████╗ ██████╗ ██╗ ██╗ ██╔════╝██╔══██╗██║██╔════╝ ╚══███╔╝██╔═══██╗██║ ██║ █████╗ ██████╔╝██║██║ ███╔╝ ██║ ██║██║ ██║ ██╔══╝ ██╔══██╗██║██║ ███╔╝ ██║ ██║██║ ██║ ███████╗██║ ██║██║╚██████╗ ███████╗╚██████╔╝╚██████╔╝ ╚══════╝╚═╝ ╚═╝╚═╝ ╚═════╝ ╚══════╝ ╚═════╝ ╚═════╝
Most ambitious problem I'm solving:
I'm building interoceptive AI. It's my bet for AGI.
External context, like email connectors, crypto/wallet connectors, or world models are limited. They only capture humans' stated preferences, which is an act we all put on, instead of revealed preferences, which is how we actually think, feel, and make decisions.
AGI needs to be human centered, and that's why we need to start by pulling context from within, tackling the latent variables.
Currently
An MBTI-style test of design taste for builders who want designs that fit their own taste.
Agent memory stored as a graph whose edges come from a fixed vocabulary. I'm betting it keeps accuracy while making lookups faster and using fewer tokens.
Lets me talk to my future self a year from now and compare what I thought and did with how I felt at the time.
Research
I corrupted part of an LLM agent's daily conversation log before it rewrote its memory each night, to test whether distorted replay helps it generalize the way the overfitted brain hypothesis says dreams do. It did not help in any of 30 runs, since the model left random text out of memory and stored changed details as fact. I submitted it to arXiv on October 14, 2026, and it is waiting to be announced.
Agent Skills
Installable as a plugin. Install once and every skill I add later reaches you on update.
/plugin marketplace add EricJujianZou/skills
Converts corporate AI writing into gen Z slang, preserving meaning. Rizz levels 1–10, from subtle word swaps up to terminal brainrot.
"brainrotify this" · "rizz this up" · "deslopify"
Removes AI-generated text indicators. Targets model leakage, corporate phrasing, and aesthetic tells. Self-critique loops and platform-specific audits.
"humanize this" · "de-AI this" · "remove the AI tells"
Source: github.com/Ericjujianzou/skills
Decision-making engine. Make the right choice, anytime, anywhere.
A Chrome extension that reads every course on Learn and answers "what's due this week?" for Waterloo students. Over 2,000 students installed it.
- Ranks what's left by due date, grade weight and effort, and skips anything you already submitted.
- Notices when a prof moves a due date and shows you the change.
A Chrome extension that adds pay per hour, applicants per opening and an ROI score to the WaterlooWorks job table.
Code, data and results for my paper Can an LLM dream? It tests whether replaying an agent's text memory with noise during a sleep phase stops it from overfitting to what it has seen.
Five pre-registered rounds testing whether frontier coding agents still need scaffolding. The scaffold won in round 5, and one test cell got caught cheating.
The harness I build my own projects with. Agents pick up tickets and open PRs, and I review them. scholarship-factory was built on it.
Play the prisoner's dilemma against five fixed strategies, then against a Q-learning agent that adapts to how you play.
Generates and manages reusable LLM prompts.
Won $5,000 at SpurHacks 2025.
A desktop pet that mirrors a live Claude Code session. She types while the agent works and puts her hand up when it needs your approval, so you can leave the terminal in the background.
Runs the voice assistant for a xiaozhi ESP32-S3 desk cat on my own machine, so the cat stops sending my audio to the seller's server.
Finds the money you qualify for and writes the applications.
- Automatically sources scholarships, grants, fellowships, internships and programs against your profile, then drafts everything for them.
- Finds the email and LinkedIn of the recruiter for your role.
- Easy onboarding and setup with your AI agent integration.
Hiring help for startups that don't have an HR person, and job hunting help for the students they hire.
- Founders describe what their last co-op struggled with, and ugmi reads every applicant's built work against that and sends back the five to interview.
- Students get same-day job listings, a named recruiter to contact at each company, and a breakdown of why they got rejected.
One dashboard for every platform you post to.
- See your live metrics update across all platforms.
- Reuse and re-purpose top-performing content across platforms.
- Automatically edit videos with AI, no extra API tokens needed.
Education
University of Waterloo
BASc Candidate, Electrical Engineering, Honours · GPA 4.0/4.0
- 2nd, SpurHacks 2025, $5,000 in funding.
- 1st, Waterloo Engineering Competition, Junior Design.
Experience
(Forward Deployed) Software Engineer · TextNow
- Owned GTM bug triage and closed 100+ cross-platform issues across web, Android and backend to GTM-ready quality, while personally shipping 50+ merged PRs across new UX features, flaky-test fixes and bug fixes.
- Automated 200+ hours of unattended code-review responses across 30+ PRs by building a more comprehensive babysitter than Cursor and Anthropic, adopted by another engineering team in 7 days.
- Raised coding agent ticket pass rate from ~77% to ~88% on a SWE-bench-style suite aimed at context rot and hallucination in fully-autonomous SDLC on production codebases, replacing single-shot cloud agents.
- Let designers verify high-fidelity UI changes without Slack back-and-forth with engineers by building a pipeline with preflight checks and Android APK builds, adopted by the head of design.
Founder · UGMI
- Built and scaled a student extension to 2,000 installs with 50% day-21 retention in 3 weeks.
- Interviewed 35 startup founders and recruiters from NVIDIA and Apple to validate the problem, then developed an ML + agentic AI system to help founders screen co-op candidates and find the best ones.
- Iterated on recruiter feedback to consider resumes instead of the contrarian GitHub-only approach, and secured a paid pilot with a seed startup for the Jan 2027 cycle.
Full-Stack AI Engineer · Revvity
- Spearheaded a biomarker analysis MVP through 4 customer-feedback cycles, pivoting from automated outlier removal to a visual classification system based on user transparency needs.
- Secured a Letter of Intent for a pilot automating 60+ reports across 20+ biomarkers by engineering an ML workflow to process 10+ years of lab data using Python, React, GPT-4o and AWS.
- Architected an LLM-as-a-Judge evaluation pipeline for 60+ agents to address user concerns on LLM transparency, cutting QA costs 7x through multimodal visual sanity checks for 2,000+ end users.
Projects
- Noisy LLM Dreaming: published a paper on noisy compaction of Opus 5.5's context, testing the theory that dreams are noisy on purpose to stop overfitting to what you've seen. Built as part of waterlooagi.com.
- Taste-Test: an MBTI-style test of design taste for builders who want designs that fit their own taste.
- Graph-Schema: agent memory stored as a graph whose edges come from a fixed vocabulary. The bet is similar accuracy with faster lookups and fewer tokens.
- Thought Sovereignty: lets me talk to my future self a year from now and cross-reference my thoughts, actions and emotions.
Skills
I'm keyword spamming for the ATS, I'm actually a genuine person.
AI/ML
Machine learning, computer vision, reinforcement learning, training, evals, transformers, vLLM, GPU
Frameworks
Next.js, Node.js, REST APIs, GraphQL, Tailwind, Three.js, TensorFlow, FastAPI, iOS, Android
Languages
Python, C/C++, TypeScript/JavaScript, Java, Go, Ruby
Cloud & MLOps
AWS, GCP, PostgreSQL, Redis, Git, system design, microservices, Docker, SQL, Linux
open with read maxxing — any word from a title works, or read 1
Contact
Any of these reach me.