ProductCamp 2026 · Ottawa

The building story behind Joe Speaking with

Joe Speaking

joespeaking.com

Joe (Beiqiao) Hu

M.Eng TIM, Carleton University Founder, Just Joe Technologies

Product Person & Vibe Engineer

Quick Survey

Let's see where we are.

Quick Survey

Let's see where we are.

01How many people here have heard of Anthropic or Claude?

02How many have used Claude?

03How many use Claude Code?

04How many use Codex?

05How many pay for Claude Max or ChatGPT Pro?

We are still very early.

AI adoption visualization — each dot is ~3.2 million people

Each dot ≈ 3.2 million people · Feb 2026

Source: @damianplayer

Press Space or Click to reveal

PRODUCTCAMP JOURNEY

ProductCamp 2025

Ottawa

Deep Research: Beyond Surface-Level Answers

Beyond the Numbers: Rethinking Data-Driven Product Decisions

This year → A real product story

ONE YEAR LATER

What Changed in One Year

As a former non-technical builder, I built my first commercial product with:

Claude Code & Codex

● Only one person

Not a prototype. A real product with real users and real revenue.

And once you enter that world, you see the whole startup process differently.

Prototype

Exciting

Experimental

Could not continue

Real Product

Users & Support

Payments & Revenue

Reliability & Trust

Analytics & Distribution

Bug Reports & Onboarding

joespeaking.com

WHAT I BUILT

Joe Speaking

joespeaking.com

An English speaking test simulator with natural conversation, real-world feedback, and a full practice lifecycle — 100× less cost, fully personalized, built by a learner for learners.

I built it for myself first — now it’s yours. My first commercial product. Real users. Real revenue.

0:00 / 0:00

Natural A.I. Conversation Recording Practice IELTS CELPIP Pay-as-you-Go

Theory-Driven Design Flash Cards Personal Collection Daily/Weekly Review

WHY I BUILT IT

01

English Was My #1 Barrier

English was the top challenge that stopped me from going abroad. I built it for myself first — then share it with others.

My English Speaking Story →

02

Research Meets Entrepreneurship

TIM master’s project needed a real product. Thanks to Hai and AGI Ventures Canada for believing in this from day one.

AGI Ventures Canada

03

Leveraging What I Already Built

100K+conversations on ChatGPT

4.4rating

500+reviews

The best way to learn is to just build things. I wanted to turn it into a real product.

PRODUCT EVOLUTION

The product itself is the evolution of how I use AI

chatgpt.com

01

IELTS Speaking Simulator GPT

100K+ conversations · 4.4 rating

ielts.hubeiqiao.com

02

Standalone Website

First independent deployment

joespeaking.com

03

Joe Speaking

Real users · Real revenue

HOW I BUILT IT

ResearchSep – Nov ’25

BuildNov ’25 – Feb ’26

LaunchFeb ’26

Evidence-Based Foundation

17

Studies

6

Theories

10

Mechanisms

Summarized from 2,877 academic records screened.

“Easy to prototype,
challenging to ship.”

🚨 Critical Bottleneck

Gemini Live API

Planned 1 week Took 5 WEEKS

REAL TRACTION

Alpha · February 5 – April 16, 2026 · $0 Ad Spend · Marketing Not Yet Started

40%

Core Flow → Paid Conversion

41

Paid Users

$350+

Net Revenue

$0

Ad Spend

LEARNING IN ACTION

300+

Simulator sessions

72.9% completion

900+

Recording sessions

4.8 avg/user

3,200+

Flashcard reviews

74.3% retention rate

1,200+

Total AI feature uses

8,000+ credits consumed

Organic Funnel

Unique Visitors 566

Logins (90%) 511

Active Users 152

Paid (27%) 41

“I don’t want my product to be popular yet — I want it to be great first.”

USAGE INTENSITY

My AI Usage

Subscriptions, daily drivers, and live token usage

Live embed • 30-day range • aiusage.hubeiqiao.com

THE SHIFT

You Can Now Skip the Middle

The old path is not gone, but it is no longer the only path.

Build quickly. Validate through action.

In uncertain territory, speed of learning through action often beats slow, data-heavy deliberation. Now I believe that even more strongly.

Connecting to Beyond the Numbers — ProductCamp 2025

The Old Path

Come up with an idea

Write a PRD

Wait for resources

Hand it off

Review

Launch

Hope it works

The New Path

Identify the problem

Build a quick version

Test it yourself

Validate directly

Iterate fast

Ship it

Know it works

🔗

This connects to what I shared last year in Beyond the Numbers: in uncertain territory, speed of learning through action beats slow, data-heavy deliberation.

WHY THIS MATTERS

The Execution Boundary Moved

This does not mean PMs need to become full-time engineers. It means the strongest PMs can now do much more of the loop themselves.

When execution gets cheaper, the bottleneck becomes judgment.

The Real Bottleneck Now

What to build Why it matters How it should work How to tell if it is good

The New PM Loop

1 Identify the problem

2 Define the workflow

3 Turn thinking into a plan

4 Build a working version

5 Verify quality

6 Iterate quickly

7 Ship to users

"The PM role now is to track both things at once: how AI is changing the way you work, and how it's changing what's possible in your product."

Cat Wu — Anthropic Research

PRODUCT JUDGMENT

My Original Thought Was Different

What I wanted to build

An AI tutor based on my practice transcripts, something that could tell me how to improve next.

Last October I tried to prototype analysis around that idea. But it was not ready yet.

I remember listening to Andrej Karpathy talk about how he learned Korean and how AI compared with human tutors. One takeaway: AI tutoring was not fully there yet in the way I wanted.

That helped me realize something important.

What I built instead

The fundamentals first.

Not the most ambitious future version
Not the smartest dream version
The inputs part. The practice loop.

Once those are strong,
more becomes possible later.

MINDSET SHIFT 2

Build from Your Own Interests

I did not build Joe Speaking only for revenue. I built it because it matched my own pain, my own curiosity, and what I wanted to learn about AI.

Joe Speaking

Even without a traditional job, I created my own job to try these workflows and feel the power of AI through something real.

The product became my classroom.

"My friends always ask me: why are you always so passionate about the things you are doing?"

Find the things you have passion for, and do them.

Not just a short project

Not just a weekend toy

Not just chasing revenue

Something you believe in enough to keep improving

Believe what you are doing.
It is a long-term game.

MINDSET SHIFT 3

You Still Need Something That Stands Out

AI makes building easier. But that means everyone can build. Differentiation still matters. You need one feature or one angle that really stands out.

Joe Speaking's Killer Feature

Real-time conversation.

Basic recording plus end-to-end infrastructure would not be enough. The thing that set it apart was also the hardest thing to build.

The Reality

My original estimate was much too optimistic. Integrating the real-time experience was far harder than the clean product story made it look from the outside.

That is where my technical knowledge was not enough. I spent more time going back and forth. But that is exactly why it matters.

You still need one thing that is hard to copy quickly.

Taste and judgment still matter.
That is where differentiation happens.

TOOLS & TECHNIQUES

What Tools I Actually Use

Not just coding tools. General agents for computer work.

Claude Code

Primary agent

Planning, coding, debugging, testing, deploying

Codex

Review & parallel work

Cross-model review, parallel audit, async tasks

As long as a task can be completed by your computer, AI is going to take it.

Build apps

Write docs

Build websites

Debug logic

Generate images by coding

Run tests

Create videos by coding

Research

Generate slide decks (like this one)

Data analyze

A very different category from a normal chatbot.

100–1,000×

More token usage than last year

HOW I WORK

My Workflow at a Glance

I shared my full workflow 3 months ago at cc.hubeiqiao.com. When I revisited it for this talk, some techniques are still helpful, some are already outdated. Here are the high-level ideas that still hold.

Mobile changes everything. Have an idea? Tell AI. Validate the result. Ship by day.

The Priority Shift

In the past, not every requirement could ship quickly. Now with AI, you can let it try. If it works, you ship it that day.

Claude Code

Main coding agent. Local terminal, deep context, multi-file edits.

Codex

Cloud tasks. Parallel branches. Background execution.

📱

Claude Mobile

Run cloud tasks or connect to local version on the go. Ideas become actions anywhere.

⚡

Ship by Day

AI tries it. If it works, ship it. Requirements that used to wait weeks now land in hours.

MY WORKFLOW

Draft the Plan First

Write what you want first. Then talk with AI to revise the plan. Multiple rounds. This is one of the best ways to demystify the black box.

Rough Draft → AI Review → Refine → Build

📋 Better breakdown

🎯 Clearer tasks

🔍 Visible assumptions

❓ Ask questions back

Use multiple models. Let Claude Code draft the plan, let Codex review it.

Full workflow guide: cc.hubeiqiao.com

Plan document with phases

Codex reviewing the plan

This talk itself followed the same path:
rough draft → plan doc → HTML slides

QUALITY

Validation Matters More Than Ever

Teach AI to verify its own work. All actions should be evidence-based. There should be enough logs.

UI

Inspect screenshots and the real interface

Logic

Run tests

Bugs

Reproduce first, then fix

Workflows

Build end-to-end tests

Verification plan

PR checks passed

The Real AI Quality Loop

1Reproduce the problem

2Write the issue clearly

3Add logs

4Verify the plan

5Implement

6Test

7Compare

8Explain

This is why real product work burns so many tokens. AI can run for hours, even days. Good usage is not just "write the code."

Lesson from Joe Speaking

I should have built the E2E harness earlier. It needs to accurately answer your questions before you trust the results.

Go through the whole process manually first. Find the cases. Then ask AI to automate them.

WORKFLOW DISCIPLINE

Work Smarter with AI

Context Management

Bigger Windows Do Not Solve Everything

Even with bigger context windows, context management still matters. Do not keep everything in one endless conversation.

Otherwise performance degrades and assumptions drift.

✂

Break work into smaller units

🔄

Open new conversations when needed

🧹

Refresh context when the task changes

Reusable Skills

Turn Repeated Workflows into Systems

When you notice a repetitive workflow, systematize it. This is one of the easiest ways to compound your own capability.

📝

Saved Prompts

Reuse what works

⚙

Scripts

Automate the steps

✅

Checklists

Catch what you miss

⚡

Slash Commands

One-tap workflows

📄

Doc Templates

Consistent structure

🔁

Reusable Patterns

Compound over time

Every repeated workflow you systematize makes you permanently faster.

DEBUGGING IS THE REAL SKILL

Hard Issues: What I Actually Do

1Let AI write a document first. List the issue clearly.

2Add logs to find the evidence.

3Let multiple models audit or analyze it.

4Search online if needed.

5Retry with a cleaner context window.

6After each failure, reflect and record.

7Ask for help: community, X, Reddit.

8Sometimes just wait for a better model.

Not coding faster. Debugging better.

When you build real products, debugging becomes the most important skill.

Incident doc

Bug investigation

Investigation report

THE PM ADVANTAGE

PMs Have an Underrated Advantage

These are exactly the things that become valuable when AI is your execution layer.

I did not "learn to code." But I became far more effective at shipping.

Even when you do not know something, you can ask AI dumb questions. The more you interact, the more you know. Learning is no longer separated from building.

What PMs Already Know

🎯 Clarify ambiguous goals

🧩 Break things down

🔄 Think in workflows

🐛 Communicate bugs clearly

✅ Define expected behavior

👁 Judge user experience

A different kind of capability shift.

BEYOND AI OUTPUT

What Building a Real Product Still Requires

It is easy to romanticize one-person AI building too much. A real product still requires a lot of things beyond AI output.

STILL REQUIRED

🧠 Product judgment

🔍 Manual QA

📊 Prioritization

⚠ Knowing when not to trust the model

❤ User empathy

⏳ Patience

🤝 Community

📣 Distribution

💪 Persistence

AI gives leverage. It does not remove responsibility.

The strongest people in this era will not just be people who know prompts.

They Will Combine

🔥High agency

⚖Good judgment

💡Clear thinking

✨Product taste

🚀Willingness to keep shipping

Not just people who know prompts. People who can ship.

FOR STARTUPS & FOR YOU

What This Means for Startups

One person can now go much deeper into the stack. Not everything forever, but enough to see the company through direct experience.

🏗 Understand constraints more directly

👥 Understand users more directly

⚖ Understand trade-offs more directly

That is one of the most valuable parts of this journey for me.

If You Remember One Thing

Please try Claude Code or Codex to build something.

Build something small. Build something useful. Build something for yourself.

The learning curve is not only about tools. It is about changing your identity from someone who waits for execution to someone who can drive it.

CLOSING

It Is Time to Build.

Not because AI is perfect. It is not.

Not because quality no longer matters. It matters even more.

Not because the hard parts disappeared. They did not.

But because the distance between product thinking and product execution has become dramatically smaller.

From a former PM without a strong technical background, to someone building a real product with real users and real revenue. Joe Speaking is my first proof. And we will see many more stories like this very soon.