Consumer Electronics / SemiconductorsSoutheast Asia
← Sticker Wall
JOURNEY · DETAIL
Xiaozao Technology Card AI Walkie-Talkie: A 3-Person Team's 100-Day Journey from Idea to Shipping for an On-Device Voice Agent
Founded: Anonymous · Xiaozao Technology
JOURNEY
Key Fields
FIELD STAMPSIndustryConsumer Electronics / Semiconductors
RegionChina
ScaleSME
ChannelOther
Origin
The team noticed that smart speakers and mobile voice assistants had low usage in mobile scenarios; users lacked a suitable device for instant voice Q&A outside the home. During an outdoor hike, members complained they could not ask AI about routes and plant information at any time, which sparked the idea of making a portable AI voice device. The team decided to use 100 days to validate the entire process from idea to mass production, avoiding the traditional hardware startup R&D cycle of up to a year.
Milestones
2025
Project Initiation Turning Point
During research, the team found that there was no on-device AI voice Q&A hardware at the 100-yuan price point on the market. Smart speakers were not portable, and mobile apps required cumbersome wake-up steps. They therefore settled on a card-sized walkie-talkie form factor, with a built-in microphone array and speaker, button-triggered voice input, and AI returning answers via the cloud. The target retail price of the whole device was 198 yuan, positioning it between children's toys and outdoor handheld radios.
2025
First Prototype Failure
The first prototype used an off-the-shelf Bluetooth audio module with a mobile app as an intermediary, but in testing, voice wake-up latency exceeded 3 seconds, and recognition rate in noisy outdoor environments was below 60%. The team found that directly applying a mobile voice solution was not viable; they had to develop their own noise reduction algorithm and switch to a dedicated AI voice chip. This raised hardware cost from an estimated 80 yuan to 120 yuan, and the project nearly stopped.
2025
Hardware Iteration Inflection Point
In 2025, the team switched to a domestic on-device AI chip, deploying speech recognition and intent understanding locally and only placing large-model inference in the cloud. In testing, the latency from button trigger to the first spoken character of the answer dropped to 0.8 seconds, and recognition rate in noisy environments rose to 85%. At this point, the team presold through online communities and received 2000 orders in one week, validating that real demand existed and securing their first 200,000 yuan in cash flow.
2025
Mass Production PMF
A 30-day crowdfunding campaign raised 470,000 yuan, with order volume exceeding 3500 units. The team collaborated with a Shenzhen contract manufacturer to deliver 3000 units in a single batch, reducing per-unit material cost to 95 yuan. User research showed that children's tutoring and outdoor hiking were the top two use cases, with a weekly active rate of 42% and a repurchase rate of 18%, confirming product-market fit.
2025
Shipping Growth
The first 5000 units all sold out, generating 990,000 yuan in revenue and a gross margin of about 25%. The team received inquiries from overseas distributors and planned to launch an English interaction firmware, while also launching a companion app to manage multiple devices. At this point, cumulative users exceeded 7000, with daily AI Q&A calls exceeding 12,000. The team remained at 3 people, maintaining operations through outsourced customer service and the contract manufacturer.
Turning Points
- Shifted from a mobile app solution to a dedicated on-device AI chip, reducing voice latency from 3 seconds to 0.8 seconds
- The presale model locked in 2000 orders before mass production, securing startup capital
- The contract manufacturer helped compress material costs, from 120 yuan in the prototype to 95 yuan in mass production
Failures & Pitfalls
- The first prototype applied a mobile voice solution, with an outdoor recognition rate below 60%
- The initial Bluetooth-connected mobile phone solution was completely abandoned due to excessive latency
- They once considered a watch form factor but shifted to a card walkie-talkie because tooling costs exceeded budget
关键成功要素
- Button-triggered voice interaction avoids continuous listening power consumption and privacy issues
- Speech recognition is weighted on-device, with only large-model inference in the cloud, reducing costs
- The presale mechanism completes market validation before production, controlling inventory risk
- Choosing domestic AI chips and a Shenzhen contract manufacturer compresses the supply chain cycle
Lessons
- AI hardware should not copy the mobile voice experience; interaction must be redesigned for the scenario
- The 100-day delivery goal forced the team to cut non-core features and prioritize subtraction
- Presale is the best way for early AI hardware to validate demand and cash flow
- Small teams outsource customer service and production, concentrating resources on algorithms and experience
Core Data
- 首批订单量：2000 units (based on public information, not independently verified)
- 众筹金额：470,000 yuan (based on public information, not independently verified)
- 累计出货量：5000 units (based on public information, not independently verified)
- 客单价：198 yuan (based on public information, not independently verified)
- 毛利率：25% (based on public information, not independently verified)
- 周活跃率：42% (based on public information, not independently verified)
- 团队规模：3 people (based on public information, not independently verified)
- 日智能调用量：12,000 times (based on public information, not independently verified)
Competitors / Peers
Similar products include the iFlytek AI Ear recording card, the not-too-distant home robot, and children's AI story machines from various brands. iFlytek AI Ear focuses on recording transcription and emphasizes office scenarios; children's story machines focus on early education content and have limited interaction depth. Xiaozao Technology's Card AI Walkie-Talkie enters through general voice Q&A, covering both children and outdoor users, but faces the risk of being crushed by large companies in chips and channels. It needs to continuously iterate scenario-based features to build barriers.