I build AI
that does the work.
I'm Evan Cao. After a decade in enterprise analytics, I now build voice agents, document pipelines and agent workflows that run inside real businesses, tuned to how each one actually works.
- PMP certified
- BI for Finning, Air Canada, Enbridge
- Founder, Sophon Studio
See it instead of reading it.
Read the transcript
Meet Evan, an AI consultant in Calgary who builds AI that does real work, inside real businesses. It starts with a decade of enterprise analytics, for companies like Finning, Air Canada and Enbridge. That's where Evan learned what useful work looks like. Out of the box, AI does things the generic way. Evan makes it do things your way. Each company's systems, terms and rules get written down as context files and reusable skills that every agent follows. Models are tested head to head on the same tasks, and each job goes to the model that earns it on quality and cost. The result is working systems. A phone agent that answers missed calls, checks live availability, and texts a booking link. An invoice pipeline where three models read every invoice, and anything they disagree on goes to a person. A complete rental platform, from online booking to the staff counter app, running a bike shop in Canmore. Plus websites, music videos and 3D, all made with the same agent workflows. Have work that needs doing? Send Evan the role or the problem, and get back a clear plan and a fixed fee.
Not demos. Working systems for real businesses.
Each one started as a job someone was doing by hand.
A phone agent that books, texts and hands off.
Answers missed calls, checks live availability and texts the caller a booking link.
How it works
- Two models, two jobs. A real-time speech model listens and talks; a cheaper reasoning model runs the tools.
- Signed connectors. The agent calls the business's own API. It never touches their database.
- Every call reviewed. A second model scores each call and flags anything the agent claimed but didn't confirm.
- Guardrails in code. It admits it's an AI, never takes card numbers and keeps spam from being transferred.
- Real-time voice
- Twilio SIP
- Node
- Postgres
Every invoice read three times. Trusted only when the reads agree.
For a property group's payables team. Three model reads per invoice; disagreements go to a person.
How it works
- Consensus reading. Three model reads run in parallel. Every money field must match to the cent. A majority isn't enough.
- Disagreement goes to a person. The approver sees exactly which fields didn't agree.
- Pluggable providers. OpenAI, Claude and Azure Document Intelligence behind one adapter. A misconfigured provider stops the app instead of faking data.
- Approvals and audit. Sequential approvals with thresholds and escalation. Every status change writes an append-only audit record.
- Next.js
- TypeScript
- OpenAI
- Claude
- Azure
- Playwright
| Field | A | B | C | Result |
|---|---|---|---|---|
| Invoice no. | INV-20418 | INV 20418 | INV-20418 | Agree |
| Date | 2026-09-14 | 2026-09-14 | 2026-09-14 | Agree |
| Subtotal | $4,180.00 | $4,180.00 | $4,180.00 | Agree |
| GST | $209.00 | $209.00 | $290.00 | Disputed |
| Total | $4,389.00 | $4,389.00 | $4,389.00 | Agree |
A rental shop that runs on three people and a lot of software.
The whole operating system behind a bike rental shop in Canmore, from booking to the staff counter.
How it works
- Booking assistant. A chat agent with eight tools answers from live availability and hands off to staff when it should.
- ID reading. A vision model reads the renter's name off their ID at check-in, with a prompt-injection guard.
- Hands-off operations. Weather-aware pricing, a 14-day demand forecast, reminders, lock codes and a nightly fleet reset.
- Express
- Postgres
- Stripe
- Twilio
- three.js
An AI production line that ships client websites.
My studio's build line: agents implement and review, I set the spec and sign off.
How it works
- One prompt, spec to deploy. A builder agent implements; a reviewer agent screenshots at desktop and phone widths and returns ship or fix.
- Locked design rules. Type scale, spacing, color and banned marketing words live in each project's context file.
- A demo factory. Over a hundred source-checked demo sites, each signed off by an independent reviewer agent.
- Claude Code
- Codex
- Playwright
- Railway
- Cloudflare
Out-of-the-box AI does it the generic way. I make it do it yours.
Your systems, terms and standards, written down once as context files and skills that every agent follows.
Break the work into deliverables
Each recurring output gets an owner, a cadence and a quality bar.
Write your context file
Systems, data sources, terminology and rules that every agent reads before it starts.
Build a skill for each workflow
A tested, written procedure: inputs, steps, checks and output format.
Route each task to the right model
Chosen from my own head-to-head evaluations on quality, cost and time.
Review and deliver
I check every output against the agreed bar before it reaches you.
- Context
- Skill
- Model
- Run
- Review
I test models before I trust them.
Same tasks, several models, graders written first. Each kind of work goes to the model that earns it.
What came out of it
| Work type | Routed to | Why |
|---|---|---|
| Writing, design, judgment calls | Top tier | Where quality is the product |
| Bounded build work with checks | Mid tier | Same pass rate, about half the cost |
| Extraction with full validators | Small tier | Cents per run when a grader catches misses |
| Anything the small model misses | Mid tier repair | All five fixed in one attempt each |
My own business already runs this way.
Agents with defined roles and handoffs, and a review step before anything goes out.
80 billion tokens. And counting.
Real numbers from my Codex profile and my Claude Code logs. Almost all of it is context the agents re-read on every turn, which is why good context files pay off.
Tokens by model. Codex split from local logs since September.
Claude Code, all sessions. Output is the part the models actually wrote.
Live sites and apps, built with the same line.
Ecommerce, booking, ticketing and marketplaces, each running for a real business.
Music, animation and 3D, made with the same agents.
Hallucination Hits, my AI music channel: Suno songs, animated entirely in code.
Claude vs ChatGPT: Number One (For Now)
An AI rap battle between two assistants. My most-watched video.
- Suno
- Code-drawn animation
- ffmpeg
Canmore in 3D
A Blender recreation of Canmore and the Bow Valley built from map data and checked against street imagery. Map data © OpenStreetMap contributors.
- Blender
- Geometry Nodes
- OpenStreetMap
Jingle ads for local trades
Thirty-second spots with an original jingle and a 3D marble animation, made for Sophon Studio's social channels. Around fifty produced so far.
- Suno
- Blender
- Code animation
Everything connects. Explore it.
My career as one map. Pick a role to see what it led to, or switch to the timeline.
Browse as a list
Pay for outcomes, not hours.
A fixed fee against agreed deliverables. You review results; I handle the agents and the fixes.
- 01
Deliverables, agreed up front
We define what done looks like: the reports, documents, systems or analyses, and the bar each has to meet.
- 02
A fixed fee
No hourly billing and no payroll. Typically around half of what you'd budget for the role.
- 03
Guaranteed
If something misses the bar we set, I keep at it until it meets it.
- 04
Your systems or mine
I can work inside your tools and access controls, or run the work on my own AI setup and hand you the output.
The fair questions.
How can one person cover a full-time role?
Most of a role's hours go to repeatable work: pulling data, drafting, formatting, checking, chasing. Agents do that part quickly once they have your context and skills. My time goes to the parts that need judgment: scoping, reviewing and fixing.
What if the work doesn't get done?
We agree on the deliverables and the quality bar before I start, and I guarantee those. If something doesn't meet the bar, I keep working on it. You're paying for the result, so the risk of it taking longer sits with me.
Do you need access to our systems?
Only if you want me working inside them. I can operate within your tools and permissions, including any AI platform you already use. Or you share the inputs, I run the work on my own setup and send back the finished output.
What about confidential data?
We decide before work starts what data can leave your environment, if any. When it can't, I work inside your systems. I follow your security and data policies, and I don't publish or reuse client material.
What kinds of work fit?
Knowledge work and IT work with clear outputs: data and BI, reporting, documentation, research, operations support, web and internal tools, and AI implementation itself. Work that needs someone physically on site is not a fit.
Is this a contract or employment?
A fixed-fee consulting engagement. I'm not on your payroll, I don't bill hourly, and you don't manage my day. You manage the deliverables.
Have work that needs doing?
Send me the role, the problem or the process. I'll come back with what I'd deliver, how I'd run it and a fixed fee.