Hackathon · Superlinked × Alibaba Cloud Qwen Hackathon · London 2026
Echo
One button on the table, and every meeting becomes something you can ask about later.
A hardware recorder that turns a meeting into searchable notes, action items and answers about what was decided.
Track Winner - Superlinked × Alibaba Cloud Qwen 2026
Won at Alibaba's office in London on 14 August 2026.
- Event
- Superlinked × Alibaba Cloud Qwen Hackathon, London
- Result
- Track Winner
- Stack
- ESP32-S3, Python, FastAPI, React, Superlinked SIE, Qwen
- Team
- Mykyta Yakivets
Why it matters
A week after a meeting, it is hard to say what was decided and who agreed to do what. Notes catch part of it, in one person's notebook. Echo is a small recorder that sits on the table: press its button and the meeting comes back as a transcript split by speaker, with a summary, the decisions and the action items. Later you can ask a question in plain words and get an answer drawn from the past meetings that match it, with those meetings named underneath.
How it works
Press once on the device, and the meeting becomes notes, decisions and owners.
Recorder
A button to record, a dial for the agent
ESP32-S3 · 16 kHz mono
REC 00:42
Question
Typed on the dashboard
Search or ask anything
Backend
Stores each meeting and its embedding
FastAPI · WebSocket · SQLite
Qwen Omni
Transcribes and splits the speakers
Alibaba Cloud · qwen3.5-omni-flash
Speaker 1: Both bugs are fixed and deployed.
Superlinked SIE
Writes the notes, finds the action items
Qwen3.6-27B or Qwen3.5-4B, by agent
- notes
- action items
- embedding
Meeting page
Summary, decisions and owners
React dashboard
Own the announcement · Speaker 1
Search
Meaning, recency and topics
Qwen3 reranker · weights 60/25/15
99% match
Qwen Max
Answers from the top five meetings
Alibaba Cloud · qwen3.8-max
Sources: Sprint planning, Payments update
Decisions we made
The dial picks the agent, the agent picks the model
Seven agents sit on the dial, from General and Standup to Fintech, Interview and Legal. Each brings its own instructions, the things to listen for and the model that writes the notes: Fintech, Interview and Legal get the larger Qwen3.6-27B, the rest the faster Qwen3.5-4B. The choice locks while recording, so a stray turn of the dial cannot change how a meeting is handled.
Trade-off. The larger model is slower. In the one run we compared, the same recording took 5.4 seconds through Fintech and 2.9 through Engineering, and only Fintech gave every action item an owner.
Superlinked first, Alibaba behind it
Each step, from notes and extraction to embeddings and reranking, is set to run on Superlinked's inference engine or on Alibaba Cloud. If a Superlinked call fails, the same step runs again on Alibaba, so an outage slows the demo down instead of stopping it. A routing page shows which model served each call and how long it took.
Trade-off. Nothing takes over when Alibaba fails, and transcription has no second option at all: Superlinked's speech endpoint failed on real audio, so who said what depends on Alibaba alone.
One voice for 25 seconds means ask again
A small microphone in a real room hears one person loudly and another faintly, and the transcriber can give the whole meeting to one speaker. So we raise the volume of the recording before sending it, and if 25 seconds or more still come back as one voice, we ask again with a prompt that says to expect several people and listen for turn-taking.
Trade-off. A long stretch of one person really talking alone pays for a second transcription, and if the retry still hears one voice, we keep the first answer.
Demo
Photos
Gallery
The full engineering is on GitHub.