Willow Frontier Pro, AssemblyAI Universal‑3.5 Pro, and On‑Device Receipt OCR

Mia: Welcome to this Bri podcast. I'm Mia. This episode comes from ProductHunt Daily, part of the Bri podcast family.

Milo: And I'm Milo.

Mia: Today we start with a wave of new voice and language models, then touch on on-device privacy, a Notion mobile app, and several developer tools that aim to make AI agents easier to build and control.

Milo: First up, two new models from Willow that are making some big claims.

Mia: Willow has put out two voice AI models called Frontier Pro and Frontier Mini.

Milo: Right. Their tagline for the Pro version is "the fastest, most accurate dictation model in the world."

Mia: They're targeting dictation, real-time transcription, and live captioning.

Milo: This is a Product Hunt launch announcement, so those speed and accuracy claims don't come with any independent benchmarks yet.

Mia: Still, it signals where competition is heating up—against players like Whisper, Deepgram, and AssemblyAI.

Milo: And speaking of AssemblyAI, they just released Universal-3.5 Pro.

Mia: It's a speech-to-text model that now supports 18 languages.

Milo: AssemblyAI calls it their most accurate STT model.

Mia: That jump to 18 languages means a developer can build voice features for a much wider user base through one API.

Milo: The launch is on Product Hunt, so we're working from the company's own description for now.

Mia: Let's move to an offline privacy story. Veryfi has added on-device receipt field detection to its Lens SDK.

Milo: This runs locally on a phone. Take a photo of a receipt and the data extraction happens right on the device—no image gets sent to a server.

Mia: That's the privacy angle. And it means the feature works without an internet connection.

Milo: The tagline is "Secure on-device extraction even if you're offline." Useful for expense management in the field, or anywhere with spotty signal.

Mia: The announcement covers both iOS and Android through the Lens SDK.

Milo: Now a story about Notion. An iOS app called "Notion Agents" appeared on Product Hunt.

Mia: It lets you chat with Notion Agents anytime. You can throw a voice note, a photo of a napkin sketch, or a text question at it.

Milo: The description we have cuts off mid-sentence, though. It promises that all requests are "handled bef...", and we just don't know the rest.

Mia: The interesting part is it decouples the AI agent from the main Notion workspace app. That could make it a quick-capture assistant while you're away from your desk.

Milo: But with only a single truncated product listing, the full capabilities and limits are still unconfirmed.

Mia: Over at Google Cloud, there's a new command-line tool for developers called agents-cli.

Milo: Its tagline: "The CLI your coding agent uses to ship agents."

Mia: The idea is one terminal command to scaffold, evaluate, and deploy an AI agent.

Milo: That covers the whole lifecycle—initial setup, testing, and launch—inside a single text-based tool.

Mia: It's listed on Product Hunt as a developer launch. Good for lowering the friction of shipping an agent, though evaluation quality in the real world is still an open question.

Milo: Another launch, called LemonLime, takes a different approach to automation.

Mia: Instead of building a workflow step by step, you describe it in a prompt. The platform self-creates the automation using the tools your team already uses.

Milo: The claim is automating existing workflows in minutes from a single prompt.

Mia: For non-technical teams, that's an appealing pitch. But the source is a launch announcement, so reliability and complexity limits aren't validated yet.

Milo: It frames the shift as moving from building automations to having them created by the AI.

Mia: On the knowledge management side, Fini has something called Knowledge Atlas.

Milo: It's positioned as a self-maintaining knowledge layer. The system ingests support tickets, writes help articles, and updates them automatically.

Mia: The description we have says it writes "he...", which is likely help center articles, but the text cuts off there.

Milo: So the idea is continuous self-improvement without manual authoring. If it works, small support teams could offload a lot of the documentation upkeep.

Mia: But again, the source is a single Product Hunt listing with an incomplete description. We can't gauge concrete accuracy yet.

Milo: Compendium is tackling a related problem for teams running multiple AI agents.

Mia: Each agent tends to end up with its own knowledge silo. Compendium is a shared workspace meant to fix that.

Milo: Companies store onboarding docs, playbooks, tables, and lessons learned. Integrations pull the data in, and the system structures it into a library.

Mia: Both human team members and AI agents can query, update, and reference the same versioned content.

Milo: The Product Hunt listing says it's "keeping your team, agents, and data on one page."

Mia: Without a shared knowledge base like this, agents operate on outdated or incomplete information, which can slow down entire workflows.

Milo: Shifting to founder tools. Eodly is a daily standup summarizer listed on Product Hunt.

Mia: It reads team activity from Slack, Telegram, Discord, GitHub, and Linear, then sends the founder one sourced summary of what shipped that day.

Milo: The tagline: "Know what your team actually shipped today."

Mia: It aims to replace a live standup with a single AI-written update. Whether the summary is accurate depends heavily on the quality of the source data from those tools.

Milo: Another listing is Orus, which describes itself as an MCP server that places a Farao trading account inside any AI chat.

Mia: That gives users chat-based access to perpetuals trading on Farao through a conversational interface.

Milo: It's an interesting move toward simplifying crypto derivatives trading. But the listing is light on details—no word yet on real-world reliability or availability.

Mia: On the content side, Bono AI claims to turn a short voice conversation into a full content strategy.

Milo: The flow: you talk for 10 minutes, and the system promises to publish everywhere.

Mia: It's pitched as a voice AI content strategist on Product Hunt. But we don't have details on output formats, quality, or which platforms it actually covers.

Milo: And finally, a hardware oddity. NanoKVM-Go is a watch-sized, serverless KVM device.

Mia: It supports 4K video, WiFi 6, and built-in AI agent MCP support. The MCP part lets an AI agent take physical control over any screen.

Milo: It's serverless, meaning it doesn't need a dedicated host computer. This effectively bridges a gap—AI software directly seeing and controlling a physical machine.

Mia: That could shift how remote infrastructure gets managed, but for now we only have the Product Hunt listing and tagline.

Mia: That's the episode. Thanks for listening.

Milo: We'll be back next time.