Kollect voice agent
Turn forms into conversations with AI
Kollect is an open source project that reimagines traditional forms as conversational experiences. Instead of filling out static inputs, users simply speak while the system listens, understands, and dynamically guides the survey in real time. Forms can also be generated by describing them in natural language, making creation faster and more intuitive.
At its core, Kollect is powered by Deepgram for real time speech recognition and voice processing. It combines modern language models for reasoning and response generation, Convex for reactive real time data management, and Next.js with TypeScript for a full stack, type safe architecture.
The goal of Kollect is educational. It was built to help developers understand how to design and implement real world AI systems that integrate voice, language models, and modern web infrastructure. More broadly, it explores how interfaces can become more human and conversational, moving beyond rigid, static web forms toward AI native user experiences.
kollect.admildomanuel.com
3 Comments
Love this approach — voice is the natural interface for forms, especially on mobile. We've been building something similar but focused on website voice assistants that can actually navigate and take actions (not just chat).
A few thoughts from our experience:
1. **Voice + Visual feedback is key** — users want to see what the AI understood, especially for data collection. Real-time transcription helps build trust.
2. **Exit ramps matter** — always give users a way to switch to text/touch. Some contexts (quiet offices, accents) make voice frustrating.
3. **Validation loops** — confirming critical info verbally before submission prevents errors. "Just to confirm, you said 555-1234, is that correct?"
The Deepgram + Convex stack is solid. Would love to see how you handle multi-step forms with conditional logic — that's where voice really shines compared to static forms.
Are you targeting any specific verticals? We found healthcare and event registration have the highest voice completion rates.
Love the concept!!! Turning forms into conversational voice flows is a strong UX shift, especially as voice models improve.
A few security and privacy questions come to mind:
• Since you process real-time voice input with Deepgram, are audio streams stored anywhere, or strictly processed transiently?
• For forms generated via natural language, are you validating the generated schema before execution to prevent injection or unsafe logic?
Voice + LLM + real-time data can become sensitive fast, especially if used for surveys collecting personal data. Making the data flow transparent, especially storage and retention, will be critical for trust.
Open source is a great move here. Security clarity will make it even stronger.
Congratulations on your launch. It looks impressive! What channels are you exploring to attract early users?