Competitor

Seven thirty in the evening.

Someone knocked on the door.

Lin Wanqing stood outside, holding a tray.

A plate of small cakes. The cream had been spread evenly, with cocoa powder dusted into the shape of a leaf on top. Several strawberries decorated the edges, all squeezed together. Two words had been piped in cream: "Thank you."

Lin Duoduo peeked out from behind her mother. "Brother Luyi! Mom made these! I helped put on the strawberries!"

Han Luyi crouched down to look at her. "Did you arrange the strawberries yourself?"

"Mm! There were five at first, but I tasted two. They were sweet."

She said it with perfect confidence, as though reporting the results of quality inspection.

Han Luyi took a piece.

"It's good. You arranged the strawberries especially well."

Lin Duoduo hugged her mother's leg and looked up. "Mom, Brother Luyi says they're tasty!"

Lin Wanqing stroked Duoduo's head, the corner of her lips curving slightly.

"I've been meaning to thank you for what happened last time." Her voice was soft, and she barely raised her eyes.

There was a faint shadow of exhaustion beneath them.

Lin Wanqing did not stay long. She left the cakes on the shoe cabinet, took Duoduo's hand, and left. Duoduo was still waving through the crack in the door. "Good night, Brother Luyi."

The door to 501 closed. He opened Sight.

Han Luyi took out his phone and opened his chat with Gu Siyue.

"A friend might need to fight a custody case. Does anyone on your side handle that?"

The reply came instantly.

"Yes. One of our partners specializes in marriage and family law. What's the situation?"

Marriage and family law. He looked it up.

"I'll tell you the specifics once I've confirmed them. The other party isn't well-off. Can the fees be flexible?"

"Our firm has legal aid slots. If she qualifies, the fees can be reduced or waived. Tell her not to panic yet."

A few seconds later, Gu Siyue sent another message.

"Also, I went over your non-compete screening list again. Out of more than a dozen people, you were the only one flagged red. That's not normal. There's another legal risk I'm verifying. Once I confirm it, I'll tell you in person."

Han Luyi stood at the door and finished the piece of cake.

Back at his computer, BugKiller v0.2 was one final step away from completion: one-click fixes.

The idea was simple. After detecting Bugs, it would feed the context to a large model to generate patches, then have BugKiller scan the patched code itself to ensure no new problems had been introduced. Detection was its strength; generation could be left to the large model. Each did its own job.

The hard part was the final step. BugKiller had to determine whether a patch had "fixed in new Bugs," but coding styles varied wildly between frameworks. React and Vue were different, and Python and Go were completely different as well. The pass rate was stuck at 93%, while Su Niannian's threshold was 95%.

Two percentage points.

He spent another three hours rewriting the template-matching logic and tried four approaches. The fourth worked.

The pass rate reached 95.2%. It was ready to submit.

The final version of BugKiller v0.2 had been packaged.

Two features had been added: Bug impact assessments now came with confidence percentages, and repair suggestions supported one-click code patch generation. He sent it to Su Niannian.

The moment he did, Su Niannian replied.

"Sent? I'll look at it right away."

Fifteen minutes later, the messages began bombarding him.

"v0.2 is so much better than v0.1. How did you get the confidence scores for the impact assessments?"

"Don't worry about how. Test it first."

Five minutes.

"Remember those five unknown Bugs we found last time? I just used v0.2 to generate patches with one click. All five passed! Han Luyi! I feel like I'm a developer now!"

She said nothing for half a minute. The product manager was doing the math.

"Han Luyi, how are we pricing this thing in the future?"

"Haven't thought about it."

"Think about it fast. A tool that can save two weeks of manpower will have no trouble selling."

"Why are you still awake?"

"I'm testing your product. Stop changing the subject. You need to take pricing seriously."

Another minute passed.

"I'll start recruiting people tomorrow. Let's get twenty beta testers first. Make an invite code."

"Alright."

Just as Han Luyi was about to shut down his computer, Su Niannian sent another message.

It was not about v0.2. It was a link.

<AI Code Review Sector Heats Up: Another Startup Secures Angel Funding>

"Take a look at this."

Han Luyi clicked it open.

It was an in-depth analysis by a tech media outlet. It focused on a startup called CodeSafe, which had just secured angel funding. The investors behind it seemed to include Dingsheng Affiliated Capital. Its founder, Zhao Wenyuan, held a PhD in Computer Science from Stanford and had spent three years at Google.

Product direction: AI-driven code review and risk assessment.

Han Luyi skimmed the rest of the article. CodeSafe had been founded last November. In five months, its team had expanded from three people to thirty. Its angel funding was enough to last at least half a year.

He opened Sight and glanced at CodeSafe's official website.

[Core approach: large models + big data (asset-heavy)]

[Public beta date: April 27]

About three weeks from now.

His energy dropped by 8 points.

At the bottom of the panel was another small line: the end of the investment chain.

[Angel funding investment chain: Dingheng Venture Capital (LP includes Prosperity Group) → personal co-investment: Wang Zhiyuan]

Wang Zhiyuan.

Han Luyi stared at those three words for two seconds. Wang Zhiyuan had also been involved in Flash's angel round—the shell company Zhang Haoran had mentioned. And now, him again.

A VP at Dingsheng, investing in startups everywhere. He was certainly busy.

He closed Sight.

Su Niannian's second message followed.

"What a coincidence, isn't it?"

"I saw it. Thirty people, money, and a team."

"Are you worried?"

"I'd be lying if I said no. But they're taking the asset-heavy large-model route, while we're going lightweight. Our approaches are different. Heavy products burn cash fast and iterate slowly; lightweight is the opposite. Their features might be more comprehensive than ours in the short term, but we have a chance in the long run."

"How big a chance?"

"I don't know. But they don't even have a product out yet, while our v0.2 worked tonight. Let's get the beta going first. We need user feedback before we can talk about anything else."

Su Niannian took a while before replying.

"Alright. Give me the invite code tomorrow. I'll recruit people first."

He shut down his computer, picked up his phone, and browsed a tech forum for a while.

<Seven Classic Pitfalls in Recommendation System Architecture Overhauls> had passed twelve thousand views. There were several new private messages—some asking for advice, some offering writing commissions, and one from an ID he recognized: a former colleague from Dingsheng GG's data team, who had joined at the same time as him.

"Hello, master. I'm a backend developer at a certain company, and we're currently running into the third pitfall you wrote about. Would it be convenient to chat privately?"

Han Luyi did not reply. This person had registered his forum account using a Dingsheng work email, which meant he still did not know who NullPointer was. But someone from Dingsheng was actively seeking NullPointer's help...

The highest-liked comment under the blog post read: "Who exactly is NullPointer? Analysis at this architectural level doesn't seem like something an ordinary developer could write."

The 404 Group was exploding too. Ma Xiaofei dropped a link: <Dingsheng Digital Announces Full Upgrade of Its Recommendation System, Covering All Short-Video and E-Commerce Apps Starting Next Week>.

Ma Xiaofei: "Damn, full rollout. Do you think it'll crash?"

Zhang Haoran: "They dare do a full rollout after only two months of system upgrades. What do you think?"

Han Luyi typed a few words: "It's not whether it will. It's when."

Ma Xiaofei: "??? You're that sure?"

Ma Xiaofei: "Then isn't Chen Bowen totally screwed?"

Han Luyi did not reply again. He recalled the D-rating from Sight's previous scan: 140 Bugs, with peak latency set to surge by 300%.

Once it went fully live, it would not last seventy-two hours.