From CTO to Solo Founder: My 2-Year Journey Building an AI Resume Agent (And 5 Hard Lessons)

---

The Spark (Origin)

In June 2023, I stepped down as CTO.

I had spent the last decade in the tech industry, managing a product with 500 million global users and 20 million DAU. The company was stable. The product was stable. But I wasn't. I was looking for something that would reignite my passion.

I didn’t rush into AI immediately. My philosophy was simple: I can learn anything. I just needed to find a problem worth solving, grind for a few years, and learn by doing.

This would be my first time building a startup entirely on my own.

Statistically, sticking to internet products gave me the highest odds of success. If a 10-year veteran couldn't make it in software, what chance did I have in a strange industry?

But what to build? 2023 was the dawn of the ChatGPT era. I had reviewed countless resumes, interviewed hundreds of candidates, and my girlfriend was an HR vet at Yahoo and Microsoft.

The idea came naturally: An AI Resume Optimizer.

The Market Check

I asked myself two questions: Is this a business? and Am I the one to do it?

1. Is it a business? I cross-verified multiple research reports. The market size is billions of dollars annually. I just needed a tiny slice of that pie.
2. The Competition? Crowded. Extremely crowded. But fragmented. There was no monopoly. If I could get seen, I had a chance.
3. The Gap? I tested every product on the market. Most were just "template sellers" or AI wrappers that hallucinated fake experience.
4. Can I do it? I had the technical skills and the aesthetic sense for resumes. I figured if I could exhaustively list every resume mistake and code rules for them, I could solve this.

Conclusion: Big market, low concentration, I can do this.

The "Perfect" Plan (Or so I thought)

I believed a successful product needed 5 pillars:

• True Demand: Verified. The industry is old, and people pay for it.
• Valid Solution: TBD (I needed to build it).
• Business Model: $100 price point seemed right to cover costs.
• Growth: To be figured out.
• Moat: To be built.

The biggest risk was the Solution. Could I actually build something that helps?

The Build: A Classic Case of Over-Engineering

Here is where I fell into the "CTO Trap."

I spent 2 months just classifying every possible resume problem. I spent 1 week reading branding books just to pick the name resylla.com.

Then came the code. Back then, AI coding tools were weak. So, I hand-coded a massive framework to support LLM calls. I built:

• Failover mechanisms (in case an API went down).
• Load balancing strategies.
• Model routing (using different models for different tasks).

"But what if I go viral?" I panicked about imaginary scaling issues. So I added Redis and Kafka. I sharded MySQL. I wrote code to support zero-downtime scaling.

The Performance Trap

LLMs were slow. My system generated 1,000+ prompts per resume. To speed it up, I implemented complex multi-threading. As any dev knows, debugging race conditions in multi-threading is a nightmare. I once spent a whole week debugging a single weird glitch.

The Result? It took me 9 months to launch the MVP.

The Reality Check

I launched. Silence. I had maybe a few dozen users (mostly students of friends). Their feedback was brutal: "Checking for errors is useless. I don't know how to WRITE the content. That's the real problem."

So, back to the code editor. I had to build the "Writer."

I used GPT-4o-mini to save costs. It was cheap, but unstable. It hallucinated constantly. I refused to accept mediocrity—I wanted a fully automated, perfect result. I spent months debugging prompts for edge cases.

Another 9 months flew by.

(Note: As I write this, I’ve switched to GPT-5.2 Thinking and Gemini 3 Pro. They are stable, reliable, and have solved the headaches that plagued me for a year.)

5 Hard Lessons I Learned

1. The "Corner Case" Trap

I obsessed over edge cases users might never see.

Lesson: For an MVP, aim for "Wow" on the core feature. Don't try to cover every boundary condition. If the product is amazing, users will forgive the occasional glitch.

2. Coding in a Silo (The "Not Invented Here" Syndrome)

I hand-coded for 1.5 years, convinced that AI-generated code was trash. Meanwhile, tools like Cursor and Claude Code were revolutionizing development. I was arrogant.

Lesson: I eventually used Codex to refactor my legacy code in 2 months. If I had embraced AI tools earlier, I would have saved a year.

3. Energy Management > Time Management

Building a complex product requires deep thought. I tried every productivity hack, but here is what works:

• 07:00 - 12:00: Deep Work (Architecture, complex logic).
• 13:00 - 18:00: Execution (Coding, bug fixing).
• 19:00+: Gym.

Environment: I go to a coffee shop for brainstorming. I only work from home for mindless tasks.

Lesson: Working from home killed my efficiency. I thought I was disciplined, but the environment matters.

4. It's a Marathon, Not a Sprint

In the beginning, I woke up at 4 AM out of excitement and coded until midnight. Result: Burnout. Insomnia. And code written at midnight often had to be rewritten the next morning.

Lesson: Now, I have strict boundaries. No coding after 8 PM. Consistency beats intensity.

5. Penny Wise, Pound Foolish

I have enough savings to live for decades. Yet, I wasted hours trying to save pennies on API costs or server fees.

Lesson: I wasted my most valuable asset—Mental Energy—to save money I didn't need to save. Focus on the product, not the penny-pinching.