We did it! π
Memory Bridge AI won 3 awards at the Activate Your Voice Hackathon by Speechmatics & The AI Collective.
π₯ Human & Communication Track
π₯ Best Use of Backboard io
π₯ Community Award
We built a proactive AI voice agent that calls people with dementia to help them remember daily tasks, track routines, analyze sentiment, and alert caregivers if distress is detected.
π₯ Demo: https://t.co/5CBJwh1WLK
π» Repo: https://t.co/g1fIsHuU08
β οΈ Hackathon prototype β some features still in development. Feedback welcome!
#VoiceAI #AI #SpeechAI #DementiaCare #Hackathon #BuildInPublic π
Hi Nicolas,
saw your post about Startup School Paris recently on twitter. I had applied a while days ago but haven't heard back, so thought I'd reach out directly.I'm an AI engineer in Paris (at https://t.co/HXdt6vYrIS, building RAG + evals for financial docs), and I ship a lot on the side:
β’ Kicky AI football shot-analysis, foundation models distilled to run on-device : demo https://t.co/qssMuXJEmi, try it out : https://t.co/HelrzmesXm
β’ Rezoume end-to-end agentic resume builder, real-time PDF, dual-LLM : demo https://t.co/IOfY5OkN1x, try it out : https://t.co/JUacKTrLmt
Would love a spot if any are left happy to share more. Either way, great lineup.
Thanks!
Abhishek
β½ This week built Kicky AI a personal football shot analyzer that runs entirely on small, local models. It's World Cup 2026. The BBC just launched a stadium-grade 3D experience for broadcast matches but most of us play on a Sunday pitch with one phone on a tripod. So for the Build Small Hackathon (Hugging Face Γ Gradio), I built the version for the rest of us. Upload a clip of a shot and it tells you: was it a goal (and when), who took it, which foot, how hard you struck it and an AI coach grades your technique with specific fixes.
How it works:
+ Zero manual labels: foundation models (SAM3 + NVIDIA LocateAnything) auto-label the footage, then I distil a fast RF-DETR segmentation model that runs in real time.
+ Pure geometry + physics turn masks into events: possession, goal detection, shot speed. MediaPipe (BlazePose) reads the shooting foot.
+ A vision-LLM coach in two modes: online NVIDIA Nemotron-Nano-12B (via llama.cpp on Modal) and a fully on-device MiniCPM-V-4.6 every model under the 32B cap, local-first.
The hardest part was the ball: 3β4 pixels wide, motion-blurred at 30fps.
Please give a like on the hugging face space if you think this is interesting <3
The whole project is the story of fighting that. Built with Hugging Face, Gradio, Modal, NVIDIA, OpenBMB, and OpenAI Codex.
https://t.co/D1BQgDq6aJ
#BuildSmall #ComputerVision #MachineLearning #HuggingFace #Football