🏆 Excited to share I got 3rd place at the @GoogleDeepMind x @vercel Hackathon NYC
I built ANSR. An AI voice host that answers every restaurant phone call. Takes orders. Makes reservations. Responds in any language.
How it works:
→ Owner uploads a photo of their menu
→ Gemini 3.1 Pro Vision extracts every item, price, and dietary tag
→ @ElevenLabs powers the voice conversation
→ @GeminiApp extracts structured data from the conversation
→ @supabase Realtime pushes orders and reservations to a live dashboard
The entire pipeline (voice in, structured data out, dashboard updated) happens with zero human input.
Built with:
→ @vercel Nextjs
→ Gemini 3.1 Pro + Flash Lite
→ @ElevenLabs Conversational AI
→@supabase Postgres + Realtime
Huge thanks to @vercel, @GoogleDeepMind, and @cerebral_valley for putting this together :)
🥉 3rd Place - Ansr
Ansr is an AI voice host that answers restaurant calls, takes orders and reservations, and converts missed calls into structured, real-time business for owners.
@ridcursion
i just launched hivememory. a shared reasoning memory for multi-agent systems.
agents share what they learn, skip redundant research, and catch contradictions automatically.
56% reuse rate. 17.5% fewer tokens. quality went up.
link below:
@garrytan for sure! shipped one I made over the weekend today.
Hivememory, shared reasoning memory for multi-agent systems. agents share structured findings, skip redundant research, and catch contradictions automatically
https://t.co/Iay3hwNPvY
I built hivememory, a shared reasoning memory layer for multi-agent systems
agents write structured claims with evidence + provenance, query what others already know before researching, and catch contradictions automatically
benchmark results with 3 parallel agents:
- 56% of queries served from memory
- 17.5% token reduction
- quality equal or better
- conflicts identified
project page: https://t.co/jOS7l3CSAY
repo: https://t.co/XKl65ZGBMj
inspired by @karpathy's LLM knowledge base. single-agent works, this makes it multi-agent.
pip install hivememory
I built hivememory, a shared reasoning memory layer for multi-agent systems
agents write structured claims with evidence + provenance, query what others already know before researching, and catch contradictions automatically
benchmark results with 3 parallel agents:
- 56% of queries served from memory
- 17.5% token reduction
- quality equal or better
- conflicts identified
project page: https://t.co/jOS7l3CSAY
repo: https://t.co/XKl65ZGBMj
inspired by @karpathy's LLM knowledge base. single-agent works, this makes it multi-agent.
pip install hivememory
Day 3: building 30 AI Agents in 30 days ✧
i think we can all agree apartment hunting is very time consuming. looking at the neighborhood, reviews, commute times, and much more.
So today, I built an agent that takes an address and pulls everything together that I'd normally spend hours digging for.
#agentsystems #automation
day 5 of building 30 ai agents in 30 days!
competitive intelligence agent
you give it a startup idea, it finds your competitors, maps them visually, compares pricing/features, and tells you what to avoid.
watch the demo below :)
#BuildingInPublic#agent
Building 30 AI agents in 30 days
Day 1: Multi-Agent Debate System
I wanted to know if AI could actually hold a real debate.
So I built 5 agents that argue, challenge assumptions, and verify claims with real sources in real-time. Turns out they can!
day 10 of building an AI agent every day!
- reads any github repo
- identifies entry points, components, relationships
- builds an interactive mind map you can explore
- generates plain english summaries
understand how a codebase works by visualizing it
Computer use is now in Claude Code.
Claude can open your apps, click through your UI, and test what it built, right from the CLI.
Now in research preview on Pro and Max plans.
@m13v_@GoogleDeepMind@vercel yeah mid-order corrections are tricky. the extraction layer picks up the final intent regardless, but handling it gracefully in the live conversation without restarting the whole order was def something i had to thoroughly test
🏆 Excited to share I got 3rd place at the @GoogleDeepMind x @vercel Hackathon NYC
I built ANSR. An AI voice host that answers every restaurant phone call. Takes orders. Makes reservations. Responds in any language.
How it works:
→ Owner uploads a photo of their menu
→ Gemini 3.1 Pro Vision extracts every item, price, and dietary tag
→ @ElevenLabs powers the voice conversation
→ @GeminiApp extracts structured data from the conversation
→ @supabase Realtime pushes orders and reservations to a live dashboard
The entire pipeline (voice in, structured data out, dashboard updated) happens with zero human input.
Built with:
→ @vercel Nextjs
→ Gemini 3.1 Pro + Flash Lite
→ @ElevenLabs Conversational AI
→@supabase Postgres + Realtime
Huge thanks to @vercel, @GoogleDeepMind, and @cerebral_valley for putting this together :)
i think the wildest finding here is how often a model tries to manipulate doesn't predict how often it succeeds. propensity ≠ efficacy
also manipulation was most effective in finance, least in health. health has ground truth that can be validated against, finance has speculation. so it's easier to get someone to make a bad investment than take a bad supplement
the domain matters more than model
As AI gets better at holding natural conversations, we need to understand how these interactions impact society.
We’re sharing new research into how AI might be misused to exploit emotions or manipulate people into making harmful choices. 🧵