AI workflows
Automate the repetitive steps: read the inbox, fill the spreadsheet, draft the reply. A person approves wherever a mistake would cost something.
Available for projects / Toronto & remote
I build AI tools
that finish the job.
Custom AI workflows, document search and assistants for small teams, including work on sensitive data that has to stay under your control. Built with the checks and records that let them run without anyone watching.
01What I build for clients
A good fit for small teams and professionals who live in documents and spreadsheets: legal, finance, operations, recruiting.
Search across your documents
Ask your contracts, policies or knowledge base a question and get an answer that cites the page it came from. Often called RAG.
Custom assistants
An assistant built for one job, connected to the tools you already use, and tested on the cases that tend to break it.
Sensitive-data automation
Client files, personal or financial records: redaction, private or local models, access logs, and nothing sent where it shouldn't go.
02Selected work
PugPort
One for all.
Every agent subscription, every account and your other computers in one chat. It picks the agent with quota left, and when one runs dry, hands the work to the next and keeps going.
Infinite agents, one ledger
Volume isn't the metric.
A shared ledger any number of AI agents can work against, run on a real, high-volume workflow for two months. What broke, and the guardrails that fixed it.
03Work with me
A demo of your workflow
Send me one task that eats your time. If it's a fit, I build a working demo on sample data in about a week. Teams and individuals welcome.
Ask for a demoBuild it properly
A fixed-scope pilot, then integration, checks and a clean handoff your team can run. You own the code and the data.
Start a projectYour data. Demos run on sample or anonymized data. Real data stays in your systems or a private cloud you control, with providers and settings that don't train on it, or local models when it can't leave the building.
04About
Engineer first. Trained in law. Building with AI.
I started in embedded control systems, then studied IP and technology law. Now I build AI tools for teams that work with documents and data, as well as my own products, like PugPort. That mix means I start from the real task, treat privacy as part of the design, test the failure cases first, and leave a record someone else can audit. Most recently: PugPort, 125k lines and 450+ test files I designed and built alone, and two months of AI agents running a live, high-volume workflow.
-
M.S. Computer Science Georgia Institute of Technology · United States Specialization in Machine Learning · Online (part-time) · In progress -
LL.M., Intellectual Property & Technology Law University of Illinois Urbana-Champaign · United States Coursework in Applied Machine Learning -
B.Eng. Electrical Engineering Dalhousie University · Canada Biomedical Engineering CertificateComputer Engineering StreamSexton Scholar
05Models I work with
01 Lead Plans, delegates, reviews
02 Specialists Build, operate, search, write
03 Local Overnight, on my own machine
04 Routing & backup Picks the tier, covers the gaps
Claude Opus 5.5
Anthropic · Lead agent
My lead agent. It plans the work, hands each piece to the right model, and reviews everything before it ships.
- Planning
- Delegating
- Final review
Runs in Claude Code
Claude Fable 5.1
Anthropic · Hard problems
For the hard ones: bugs nobody can pin down, design calls that touch everything, questions without an obvious answer.
- Root causes
- Architecture
- Second opinions
Runs in Claude Code
GPT-6 Astra
OpenAI · Complex builds
Where real applications get built. Give it a scoped brief and a repo, and it comes back with a tested change.
- Features
- Refactors
- Test repair
Runs in Codex
GPT-6.1 Terra
OpenAI · Computer use
My hands on a screen. When a job means clicking through an app or driving my other computer, Terra does it.
- GUI automation
- Remote machines
- Visual checks
Runs in Codex
Grok 4.7
xAI · Wider answers
When I want the full range of answers rather than the safest one. Grok searches wide and speaks plainly.
- Open questions
- Live search
- Blunt takes
Runs in Grok CLI
Gemini 4 Argon
Google · Writing · GPUs
My writer and my training rig. Gemini drafts long-form prose, and Colab GPUs run my fine-tuning jobs.
- Long-form writing
- Rented GPUs
- Fine-tuning
Runs in Gemini app · Colab
Qwen3.8-27B
Qwen · Overnight runs
Runs on my own machine. Long jobs and overnight batches go here: no quota, no meter, and nothing leaves the desk.
- Overnight batches
- Long runs
- Private data
Runs locally
Recursive Flint
Recursive · Night-mode coding
A fast local coder for night mode. It takes small coding tasks and cross-checks the day's changes while I sleep.
- Small coding tasks
- Cross-checks
- Night mode
Runs locally
Laya
My own fine-tune · Task triage
A model I fine-tuned myself. It reads each request in under a second, grades how hard it is, and picks the model and effort level.
- Task grading
- Model choice
- Effort level
Runs in PugPort
DeepSeek V4
DeepSeek · Low-cost backup
The low-cost failover. When another provider hits its limit, the work moves here. It's also my pick for reverse engineering.
- Overflow work
- Reverse engineering
- Bulk jobs
Runs via API
06Shipping lately
JulAugSepOct
07Notes
08Contact
Tell me what eats your week.
A few sentences are enough. I reply within two working days.
- Emailwang@yutong.wiki
- LinkedInYutong Wang
- GitHubYutong43