naly was an AI companion speaker for freelancers and gig workers who work alone — developers, designers, writers — basically Alexa with a personality, memory, and the willingness to take your side instead of staying politely neutral. The bet was that people working solo are lonely, want a friend rather than a neutral assistant, and would pay for a device that talks to them about their work. I co-founded it in January 2024 and ran the whole hardware and embedded side. We raised $65K+ in non-equity funding, won the Grand Prize at the SKKU Creative ICT Competition, shipped working prototypes, and then pivoted after 10 months once it was clear the hardware economics and timelines wouldn't fit our runway.
naly MVP — wake word, conversation, and on-device memory in action.
naly — from PCB design through DIY reflow oven to the Shenzhen production line.
Customer discovery, before any firmware
We validated demand before building the real product: one rough prototype video, posted with zero ad spend, went viral to 25,000+ people on Twitter in 24 hours and pulled in $6K+ in pre-orders. Then I called the buyers and interviewed them. Almost all of them turned out to be people who work alone — freelance developers, designers, writers, founders-in-the-making — lonely in their work, short on anyone to talk shop with, and hungry for a companion that understood what they did. That's who we built for, and it shaped the wake-word behavior, the voice persona, and the things we chose not to build. It stuck as a habit I still keep: talk to your real buyers before you trust any PRD.
Architecture
- SoC: Qualcomm QCS405 (via Thundercomm, a Qualcomm subsidiary).
- Embedded OS: a custom Linux image built with the Yocto Project. I owned the Board Support Package: kernel recipes, device tree, audio drivers, system services.
- Audio stack: kernel-level tuning of the DSP path, because low-latency wake-word plus full-duplex voice was the whole UX.
- Hardware: multi-layer PCB (DDR, PMIC, mic array, speaker amp) designed in EasyEDA and KiCad.
- Edge AI: an ONNX wake-word model running on-device, small enough to run cheaply while staying accurate.
- IoT backend: AWS IoT Core for device registration, MQTT telemetry, and OTA firmware updates.
The hardest moment
An on-site debugging trip to Shenzhen. Samples coming off the line had an intermittent audio glitch, clicks and dropouts that showed up on the manufactured boards but never on our dev units. I spent a week in the factory with an oscilloscope and a logic analyzer and tracked it down to a grounding and layout issue, made worse by a slightly different crystal tolerance on the manufactured BOM. We reworked the layout and confirmed it on the next run. Standing right next to a problem, on a deadline, in a language I barely speak, teaches you something no amount of software-only work does.
What it taught me
I went from knowing nothing about hardware to shipping a custom Linux OS, a multi-layer PCB, and a factory debug in under 10 months. If you want proof I learn fast and ship fast, that's it. It also taught me to kill projects on time. Admitting naly wasn't going to make it was brutal, but that pivot is exactly what freed us up to build mefriend.