The newly launched AI notetaker Ellis faces heat from the developer community over privacy issues and its 'confident ignorance' during crosstalk.
Taking manual notes during an in-person meeting makes you look like a corporate slave, but buying a $500 AI pin that ends up in your drawer makes you look like a total sucker. Enter Ellis, a consumer-first AI notetaker designed to run directly on your iPhone and Apple Watch.
Sounds like a solid pitch, right? But the dev community on Product Hunt didn't let creator Robin off that easy. Within hours of launch, they started poking massive holes in its armor, specifically targeting privacy concerns and how the app handles "crosstalk" (spoiler: it gaslights you).
According to the creator, Ellis is built for individuals, not enterprise organizations. It’s meant to record and summarize any real-world interaction—from a casual coffee chat or a sales meeting to sensitive personal moments like doctor visits or therapy sessions.
Under the hood, the workflow is straightforward:
It sounds great on paper, but seasoned developers immediately spotted some massive edge cases.
One sharp-eyed user raised a critical issue that plague most diarization APIs today: overlapping speech.
In fast-paced, real-world brainstorming sessions, people constantly talk over each other. Instead of marking these overlaps as "crosstalk," Ellis’s underlying engine (AssemblyAI) simply picks a winner (whoever is louder) and completely discards the other speaker's words.
This leads to a clean-looking transcript that is, in fact, completely wrong. As the user pointed out: "A confidently wrong transcript is worse than one that admits a gap."
When asked if Ellis could at least drop a "crosstalk" warning marker, the creator responded with a classic developer deflection: "Is this a problem worth solving using tech or meeting cadence?" In other words: "Is it my code that's bad, or do you guys just not know how to take turns talking?" Classic.
Ellis's marketing suggests using the app for highly sensitive scenarios like therapy or medical consultations. However, when pressed on whether the voice processing happens locally on-device, the creator admitted it all gets uploaded to a server.
Even though the recordings are allegedly deleted immediately after transcription, security-minded folks are raising eyebrows. If you're recording third parties who haven't consented to having their voices processed on a random cloud server, you are stepping into a legal minefield.
On the bright side, the community praised Ellis for not trying to sell a proprietary hardware gimmick. You don't need fancy AR Translation Glasses with ChatGPT or an expensive dedicated clip-on microphone. It works on the devices you already own, which keeps it practical.
Building an AI wrapper is easy; making it robust against the chaos of real human behavior is where the real engineering lies. For those looking to build their own AI micro-SaaS, here are two major takeaways:
Stay safe, and may your transcripts never gaslight you!
Source: Product Hunt
