Dingsheng's Outsourcing

The next morning, the first thing Han Luyi did after waking up was turn on his computer and reopen that outsourcing request.

He hadn't made a decision last night, but after sleeping on it, he had figured it out.

The technical overview attached to the request was only three pages long. It had been anonymized, with the company and project names removed, but to Han Luyi, it might as well not have been redacted at all.

He could match every description to a specific code file.

"The core recommendation engine recently underwent an architectural adjustment, replacing the original dynamic weight allocation module with a rule-based static engine."

Han Luyi's expression turned rather strange when he saw that sentence.

Dynamic weight allocation—that was the core scheduling module he had spent eight months designing, the heart of the entire recommendation system. Its brilliance lay in its ability to automatically adjust recommendation weights based on users' real-time behavior. Every click, every swipe, every pause was factored into the calculation.

This wasn't something that worked the moment it was written. He had coded and tuned it at the same time, running hundreds of millions of user data records and grinding through the parameters one by one.

Then Chen Bowen's team had dismantled it and replaced it with a static rule engine.

Han Luyi understood why. The dynamic weight module was indeed complicated. Without him there, no one else on the team could fully understand that logic. Chen Bowen's choice had been simple: if he couldn't understand it, replace it with something he could.

The technical overview also listed a series of symptoms: response latency spiking during peak periods, declining recommendation accuracy, continuously falling user retention, and A/B test results consistently worse than the old version.

To put it plainly: after the replacement, everything had gone to hell—and it was getting worse.

Han Luyi leaned back in his chair and opened Sight.

His energy was full: 250/250. He focused his attention on the three pages of technical documentation on the screen.

The panel exploded.

This wasn't an ordinary pop-up. The entire display was covered by a translucent red warning, like a wall on fire.

[Target System: Dingsheng Digital Technology · Core Recommendation Engine]

[Overall Stability Rating: Extremely Dangerous]

[Estimated Time to Total Collapse: Within 72 Hours of Launch]

One hundred and forty Bugs, dark red and light red, densely covering the screen like a critical illness notice.

He only glanced at the first three severe Bugs before he no longer wanted to look. Memory leak—the system would choke itself to death in three days. Broken data pipeline—one-quarter of user data was quietly being lost. Concurrent lockup—during peak traffic, there was a twenty percent chance of an outright cascade failure.

Three ticking bombs. Any one of them blowing up would be enough to make the trending searches.

He had spent three years taking this recommendation system from barely usable to genuinely effective. Every parameter, every weight, every version iteration had been typed out by him line by line.

Chen Bowen had only taken over half a month ago.

He remembered the line on Chen Bowen's panel on the first day he was laid off—

[WARNING: This target has 3 severe Bugs. Click to expand. (Current level insufficient)]

At the time, he had been Lv. 1 and couldn't see the details.

One of those three severe Bugs was probably this. Chen Bowen had picked his fruit, but he couldn't chew it.

Of course Han Luyi could fix it. He had written this code himself. He didn't even need Sight; for some of the Bugs, one look was enough for him to know exactly which line to change.

He closed Sight.

The screen returned to the normal outsourcing request page. In the upper-right corner were two buttons: [Accept Request] and [Ignore].

Han Luyi's finger hovered over the mouse.

Accept it? Save Chen Bowen? Am I an idiot?

Then he clicked [Ignore].

He wasn't taking it.

He closed the outsourcing page and took a sip of the coffee on his desk, already cold.

Not taking the job was one thing, but keeping certain things bottled up was a waste too.

He opened the forum and, under the name NullPointer, spent forty minutes writing a purely technical post titled <Seven Classic Pitfalls in Recommendation System Architecture Refactoring>.

Those were all pitfalls Chen Bowen was currently stumbling into. But Han Luyi knew that, at Chen Bowen's level, even if he saw the post, he wouldn't understand it. When the system really collapsed, and this post was dug up again, it was hard to imagine how spectacular Chen Bowen's expression would be.

After posting it, he closed the forum.

His mind began to turn.

These 140 Bugs, the hundred-plus Bugs he had fixed through freelance jobs over the past two weeks, and before that, the thousands of Bugs he had handled over five years.

They looked different, but the patterns were all the same.

The same patterns appeared over and over in different projects. Every time he fixed a Bug, he was essentially doing two things: recognizing the pattern and matching it to a solution.

Wasn't that exactly the kind of work AI could be fed?

He set down his cup, took out his phone, and found the line in his notes—

"Code review tool, connecting code and real-world scenarios, BugKiller?"

Organize every Bug pattern Sight had seen—the characteristics, trigger conditions, and repair paths—then feed them to AI. Teach it to "look at code and know what problems will arise after launch." The AI didn't need Sight. It only needed to learn what Sight had seen.

Like teaching AI to play chess—you didn't need to teach it how to think. You only needed to let it see enough games.

Han Luyi opened his computer, created a new folder named "BugKiller," then created a new document: Bug Pattern Database v0.1.

Then he started writing. He wrote quickly. Every pitfall he had encountered over the past five years surged out of his memory and automatically fell into place by category: memory-related, concurrency-related, data flow-related, interface-related. Under each category were trigger conditions, typical symptoms, and repair paths.

It was as if he were compiling a Complete Guide to Bugs for a novice who knew nothing at all.

After writing for two hours, he stopped.

It wasn't enough. One person's experience couldn't support an entire database. He needed more samples and more systematic classification.

He needed a product manager.

Han Luyi picked up his phone and opened his WeChat chat with Su Niannian.

"Niannian, I've got an idea about the AI code review direction we discussed last time. Are you free tonight? Let's have dinner and talk."

Less than thirty seconds after he sent it, Su Niannian replied: "Tonight?! What time?"

"Seven."

"No problem! I get off work at six-thirty anyway. What's the idea? Tell me a little first?"

"I'll tell you when we get there."

The "The other party is typing..." message at the top of the chat flashed several times. In the end, what came through wasn't another question, but:

"Then I'll pick a restaurant first. What do you feel like eating?"

"Anything. I'm not picky."

"Okay, I'll decide."

Han Luyi found it a little funny. He had been the one to invite her, yet after three sentences, she had taken control.

Still, that was normal. Su Niannian was a product manager. Seizing control of requirements was instinctive for her.

Han Luyi put down his phone and buried himself in the document again. When he looked up, it was already four-thirty in the afternoon.

There were still two and a half hours until seven.

He turned to look at the BugKiller folder he had just created on the screen. After thinking it over, he didn't continue writing. When he talked with Su Niannian tonight, he needed his ideas to be alive in his head, not stiffened into words on a page.

He shut down his computer, got up to take a shower, and changed into a clean shirt.

As he left, Lin Duoduo's voice drifted over from next door as she told Lin Wanqing about kindergarten.

Han Luyi closed the door behind him and headed toward the subway station.

The evening breeze was neither cool nor warm. Haicheng in mid-March sat precisely in the crack between two seasons.