Demo Day S2

This is season 2 of Demo Day, which we run once a month.

Michael takes the first slot

I kicked off the second round of The Stage at SQ Collective. I shared upfront that I made little progress on Kai since my last demo.

For those who are new, I vibe-coded Kai, an AI community manager, and Rex, a back-end automation system. They are both meant to augment the very traditional business of SQ Collective using AI.

Both of them work, and I thought they were brilliant ideas. But nobody's particularly reaching for them though. No customers are using Kai, and no staff are using Rex.

Writing code has become very easy, but it does not mean you get to skip the very human problems: Why are you doing this, and what exact problem are you trying to solve for your audience?

"There's a lot that goes into designing products that people actually want to use," he says. "If you don't figure out why people will use these things, it doesn't mean anything."

Tanmay's progress update from S1: Interview Prep Gets Smarter

Tanmay's app, Final Round, was here last Demo Day when the Mac version was still in review. It's out now.

The improvements since then came mostly from feedback in this room.

• Russell flagged that questions were truncated and users weren't given time to think.

• Tony found sentiment scoring was miscategorising answers.

Tanmay fixed the timing flow and trained a Core ML model on Apple's NPU to handle sentiment - classic ML, on-device, trained on labelled examples from this community's feedback. He also built a RAG pipeline pulling real interview questions from Trustpilot and open-source communities to give the evaluation layer actual signal about what correct answers look like by job category.

The live demo runs through question generation, a timed response flow, transcript output, and a confidence breakdown at the end.

"The sentiment breakdown is more accurate now. This will improve with time as I get more data."

Stefan and OpenClaw: The Autonomous Loop

Stefan is a non-technical entrepreneur with multiple businesses in Germany.

He came into SQ Collective 4 weeks ago "learning about AI".

Fast forward today, he has built himself a personal AI agent that runs on OpenClaw and works around the clock - watching every signal in his environment: WhatsApp, LinkedIn, Teams transcripts, emails, web monitoring. It enriches each signal against a database of people and projects he calls "cases," then starts working through them without being asked.

In his demo, he types a message to his agent in Telegram, it retrieves context, surfaces a decision request, suggests next steps. A button press approves it. For routine decisions, Stefan has given the agent authority to decide at 91% confidence without him. Budget overruns still escalate while everything else gets done without him.

His hardest design lesson: breaking tasks into steps with different authorities - understand context, create approach, identify tools, create a decision milestone - produces much better quality than asking the agent to "just fulfil it completely." Someone from the room asks how much of this he built outside OpenClaw.

"My last message to Claude Code on my laptop was: I heard about OpenClaw, I want to try it. Everything after that, I just built it using Claude through Telegram. I don't know the details, but it works."

Many of the attendees have reached out to ask him how he did it.

Zhuliang and Roblox Coding Agent

Zhuliang's startup, Lemonade, is a live coding platform for building Roblox games using AI agents. Think Cursor, but purpose-built for Roblox Studio. His co-founder is in the US. They're currently number 11 on OpenRouter.

The demo: an audience member requests a FarmVille-style game. Zhuliang fires the prompt at Lemonade, an agent in the cloud generates a plan, and inside Roblox Studio you watch it create scripts live. The first pass produces two agents that don't do anything. He prompts again - "it doesn't work, make it interactive" - and gets an interaction trigger on the second attempt.

Complex prompts confuse the agent. Simple ones work. But the user base is a surprise: "It's more for kids, but we also have users who are 30, 40, 50 years old, trying to build their favourite game from childhood."

Lukas: Forecasting Without the Guesswork

Lukas presents EOMER, built around tabular foundation models - these are pre-trained models for structured numerical data. First production use case: time series forecasting.

The problem: most enterprise forecasting is either gut instinct from a 30-year veteran, Excel, or SAP tools that nobody loves but everyone already pays for. Pre-trained foundation models beat these on complex tasks.

The demo works: upload a CSV, configure covariates and forecast horizon, submit. In seconds you get confidence intervals rather than point estimates. "The avocado sales at FairPrice in Katong will be between 13 and 21 next week" - not a single number that looks precise but might be wrong. The distribution makes planning more robust.

Regression and classification pillars are still being built. Time series is live. They've found a lean compute partnership to avoid AWS costs, and can fine-tune models for specific clients and markets.

Some attendees in the FMCG space gave him useful feedback after the session, and also offered to provide some data for him to test out.

A Robot in the Room

Siddharth demos robot teleoperation live - a manipulator arm in his house responding to controls with around 50ms latency. He's spent seven months scouring for hardware across San Francisco, New York, and China. Three types of robots, he runs the models and teleoperation, helps enterprises deploy fast.

Target markets: hospitality, office spaces, concierge robotics, chip assembly lines. VR teleoperation across countries under 100ms is coming - he thinks he can show that in April.

He also mentioned that there might be a robot coming to help Michael push in chairs after an event. wink

Eugene: Vibe-Coded Into a Partnership

Eugene signed up the day before, has no slides, and opens with: "I haven't touched any code until eight months ago. I started vibe coding two weeks ago."

In those two weeks he built a construction materials data aggregator for architectural firms. The problem: architects choosing sustainable materials have to manually compare products across dozens of supplier websites. He talked to George Bengtsson, co-founder of Edge Impact, over two lunches, and has now been invited to Australia to sign a partnership with Grimshaw, one of the largest architectural firms there.

The backend is in Supabase: 14 tables, APIs from multiple sources, an ETL pipeline that normalises material data, and a scoring system built on a sustainability methodology implemented entirely through Claude Code conversations. Architects compare products across five pillars and drill into why something scored four out of five on circularity.

"I'm better at connecting people and raising money," he says. "Now I'm also learning to vibe code."

Luke: The Agent Marketplace

Luke Hsieh, co-founder of Agently, is building infrastructure for the moment when hiring agents becomes as normal as hiring contractors. Two sides: AIXYZ, a payment SDK so agent builders can monetise per call (stablecoins now, Stripe coming), and a discovery layer so businesses can route tasks to the right agent by cost, speed, and reliability.

The model is Uber or Upwork for agents. Compute stays with the builder. Agently routes and handles payment. He demos a user telling their own agent to "use Agently to search X posts about topic Y" - the platform finds the best match, routes the task, bills automatically. A month in. The ecosystem is thin, which he says upfront.

Closing: Farhad shows vibe music and the Power of WebGPU rendering

Then Farhad closes with something nobody expected. His background is graphics programming. He opens by showing a full GPU video editor built in the browser: split, cut, mask, blend, export, ML models running client-side, zero subscriptions. He thinks we are sitting on a Figma moment - like when Figma saw WebGL in 2011 and realised they could build a native-quality design tool in the browser. (Check it out here: https://www.masterselects.com/ - this is 100% open-sourced, vibe coded by a movie director, FREE, and does AI/ML/GPU operations ALL within the browser).

Then he opens Strudel, a programmatic music library, pastes a prompt into Claude asking for "a cool deep techno Berlin style, ambient, slow bass, dark vibes," and plays the result through the speakers. The room shouts requests. Saxophone. Deeper bass. Every four beats. He adds them, ties the output to a WebGL visualiser mapped to the beat, and the session ends with generative music and live shader visuals.

"I found Claude does a much better job understanding Strudel compared to Gemini or ChatGPT," he says. Then: "Try Strudel. It's pretty cool."

---

Michael started by confessing his AI employees have no real users. By the end, Tanmay had shown Core ML models trained on feedback from this room. Stefan's agent was approving decisions on his behalf, live. Eugene showed up with no slides and a two-week build that landed him a trip to Australia. And Farhad had the whole room rocking their headas to a techno track together. These Fridays have their own kind of energy.

Missed out last week? Don't worry, these conversations happen every Friday at SQ Collective.

Usually over laptops. Sometimes over pizza.

You're welcome to join the next one.

Watch our recording of the session here: https://youtu.be/62NhmfutgOY