How an approach to the “atmosphere of the atmosphere” to peoplepark tripled the arrival growth and supported a dam of new products and features in just a few weeks

Personk, launched in June 2024 by Mainfunc, was initially focused on the search for AI. But despite reaching 5 million impressive users, the company has moved away from this initial product to Super Agent, which, instead of following a static sequence of steps as in traditional research, chooses the best tools or sub-agents for work, my results and adjust in real time.

Launched on April 2, Super Agent is propelled by Claude d’Anthropic and can condense an afternoon of white necklace work in 5 minutes, says Zhu. For example, he can make calls, download, check the facts, produce podcasts, write documents, carry out in -depth research and collect the spreadsheets and the slides.

“We always see it as a kind of research, but it is more technically advanced,” said Zhu, who has more than 20 years of experience in research at Google and Baidu.

The company has aggressively added more and more features in the last four months; Here is an overview of its deployments and its milestones:

• April 11: reaches $ 10 million arr 9 days old after the launch of super agent
• April 22: introduces AI slides (with hundreds of models)
• April 28: deployed a super personalized agent with adaptive personalities
• May 2: Hit $ 22 million Arr, exactly a month after the launch
• May 8: launched intermediary sheets which create complete spreadsheets in one click
• May 15: introduces a fully agentic download agent and an AI reader that manages and stores files
• May 19: hit $ 36 million round
• May 22: Deployed AI which can make phone calls
• June 4: presented an AI secretary who manages Gmail, Calendars and Google Drive
• June 10: Deployed an AI browser and an MCP store with extended navigation capacities and a tool market
• June 18: introduces AI documents for the creation and management of documents
• June 25: Introduced Design Studio with “canva -shaped” capacities for the creation of visual content
• July 10: deployed pods ai to create podcasts with simple prompts
• July 17: Presentation of advanced publishing features for IA slides
• July 31: deployed of slides AI 2.0
• August 1: Multi-agent orchestration that can produce up to 10 agents simultaneously

PEOSPLK also warms the AI agent’s space with a friendly competition. After Openai announced his agent Chatgpt in mid-July, Perspark carried out a comparative analysis and is “very confident” in his ability to outperform the rival. To drive this point to the house, the company launched a “confrontation of AI side by side of a million dollars”, challenging users to chase the cases where other platforms surpass the super agent of Perspark.

In the first round, users were responsible for building a 12 -page financial slide using agent peoplepack and chatgpt; Users have identified 429 cases where the latter surpassed the first, each winning $ 100 for their efforts.

In turn 2 (which ended on Monday August 4), people park increased the bet to $ 200 by victory and opened the competition to any AI tool as an adversary. Users were challenged to use exactly the same prompt to create slides on people park and their chosen AI tool, then download them to Gemini for evaluation.

“Do not try to start a drama here – simply enthusiastic by the way to the way to the ecosystem of the AI agent came,” posted the company on X. “This shows that we are pushing all the limits in the right direction.”

Some user reactions:

How the native team of the AI of people park vibes

The secret of people park is his lean and native team of 20 people and engineering philosophy of “less control, more tools”. Zhu explained that more than 80% of his code is written by AI, which is not an atmosphere coding in itself, “because the ambient coding indicates that you never look at the code.” PEOPLK rather has a “very rigid” code examination process to help guarantee the quality of their code base.

“We only need a very small a-native team to operate in superhero mode, as The Avengers“Said Zhu, who said that they would gradually add the team members if necessary.” AI coding and AI workflow are so powerful, it’s a magnifying glass. “

Today’s business teams must be reorganized “totally differently,” he said. He managed 1,000 teams with different levels of management and given how office policy can introduce friction.

The team of people park, on the other hand, communicates “very transparent” and productivity is “super high”. “Everyone is working on a product that can be shipped,” said Zhu. “I believe it will be the standard for waiting, because AI really helps more and more people to do their jobs better.”

He also stressed the importance of immersing you in your own product. From the designers themselves to the marketing team, “we actually eat our own dog food. We are our own consumer of products. This is how we will continue to improve the experience. ”

Inside the great flagship agent of Perspark

Zhu noted that when the perplexity was launched in December 2022, this triggered the excitation as to the potential of the AI to transform research. However, he followed rigid workflows, with platforms in front:

• Analyze requests and extend keywords;
• Recover the web results of the best;
• Rerank / Summary for a final response.

This was enough for basic things, but “crumbled” in more complex scenarios such as technical comparisons, in-depth research and multi-stage and multi-factor purchases. “In essence, it was like trying to sail in a labyrinth with only fixed turns,” said Zhu.

Personk has built its search engine on this same type of foundation, superimposing incremental improvements, including specialized data sources, parallel research of a more in -depth survey on complex requests and cross -checking of asynchronous agents to check the declarations that are too complex for “rapid manipulation and on the fly”. But they realized that they were always “chained” by fixed and predefined workflows, Zhu reported.

Super Agent uses nine models of large language (LLMS) of different size and differently specialized in an agent mixture system (MOE). The models decompose the tasks into steps, by delegating according to the specialty and the force, then are transmitted. Super Agent is also equipped with more than 80 tools (sub-agents that can generate Python code to those who can make telephone calls independently) and more than 10 sets of data organized from the web, partners and benchmarks.

Persvark gives tasks to Claude, Openai, Google Gemini, Deepseek., Ai’s Grok 4 and others: “Then we let everyone produce their production, and we have a model of aggregator to examine the results and analyze what process is the most profitable,” said Zhu. “In this way, we improve precision, reduce hallucinations.”

The company also refines its own border model. However, they are not too aggressive as to the creation of advanced systems like Deepseek V3 or V4, said Zhu. The objective is to ensure that the model performs low -level but heavy work.

“We are not trying to push the limit of the frontier model,” he said. “We are trying to reduce the cost and latency, because many proprietary models are too large, too slow and too expensive for many relatively simple tasks.”

As for the tendency to cod the atmosphere, the objective of people park is to allow everyone to experiment, even for non-programmers where the concept can be a bit “too distant”.

“Many people think”, atmosphere coding, I heard about it, it seems cool, but I don’t know the integrated developer environment (IDE), I don’t know the code, “said Zhu.” Using people park, people can really vibrate. ”

Daily information on business use cases with VB daily

If you want to impress your boss, VB Daily has covered you. We give you the interior scoop on what companies do with a generative AI, from regulatory changes to practical deployments, so that you can share information for a maximum return on investment.

Read our privacy policy

Thank you for subscribing. Discover more VB newsletters here.

An error occurred.