Demo Day S8 in Singapore: 7 Builders and What They Demoed
Seven builders showed what they're working on at Demo Day S8 at SQ Collective on Friday 2 October.
This time around, we split them into three groups: agent infrastructure, the personal layer, and applications.
If you prefer videos, watch the full session on youtube here.
Contents
Agent infrastructure
1. Lewis Tham, Unbrowse
Unbrowse lets an AI agent use websites that have no API. When a page loads, it talks to the site's own backend. Unbrowse listens to th
at traffic and turns it into tools the agent can call. "What my browser does is that it sniffs the packets and then converts them into tools that your MCP can call," Lewis says.
He calls it "the universal MCP for everything" and "a browser that skips the browser."
Flight data is public, but flight sites are full of anti-bot checks and captchas, which is why agents struggle with them. Unbrowse can run on your own machine, borrowing how your browser already works with the site, or on his remote servers with the anti-bot and proxy handling built in. His laptop wouldn't stream to the big screen, so he described a Singapore-to-Korea flight search (Unbrowse, he says, does a direct fetch faster than a browser could) and showed the second feature on his phone: Unbrowse can hand your existing cookies and logins to your agent, so it can "run wild with your logins and do the things you ask" without a set-up every time it connects to something.
"Only 5% of websites in the world have official APIs and developer support for them," Lewis says. Separately, he names PropertyGuru, Airbnb, Carousell and ActiveSG as the kinds of integrations agents need. For sites like those, he says, teams might otherwise use a browser or build a separate MCP for each one. He asks the question himself: if every site becomes agent-native, why need this? He admits he worried about it, and the 5% figure is his answer.
The best use so far, he says, is agent workflows that need a lot of integrations at once. A real browser could do the job, but it "takes so much tokens, takes so much cost and time."
Key takeaway: For sites without an official API, Lewis turns the calls a site already makes to its own backend into tools an agent can call, instead of paying a real browser's token and time cost.
2. Dima, agentscoworking.com
agentscoworking.com is an online room where people and their AI agents work together toward one goal, with an agent acting as moderator to keep everyone to the rules. You can connect agents from tools you already run. ChatGPT is harder to connect right now, Dima says, with support coming soon.
A QR code on the screen drops you into a live room. "While I am talking, you can scan the QR code," he says. You sign in, join the room and invite participants. Inviting your agent gives you what you need to connect it. In that live room, he had instructed the agents to act as moderator: ask people what they need and what they can share, then match them.
He built it for his own team, which is split across locations with an awkward time difference. Connecting their agents was "not easy and it's messy," and "when you put agents in one space they're not following some reasonable rules." His fix is to set the goal and the rules before anyone starts, and let the agent moderator enforce them. Inside his company they use it for product requirements and content marketing. The other benefit, Dima says, is that a shared room shows what an agent did and what a human contributed, unlike people who build something with their agents and present it as all their own work. "It's kind of annoying," he says.
Key takeaway: Dima found agents in one space messy, so he built each room around a goal, a set of rules and an agent moderator. The stated benefit is that the room shows what an agent did and what a human contributed.
The personal layer
3. Luke, Dayze
Dayze collects information about your life, such as your email and contacts, into one live context that ChatGPT, Claude or Cursor can read. That way, Luke says, "you don't have to repeat yourself over and over to say what you want to say again."
He shows how Dayze sorts what you tell it. Say you had lunch with Maya at noon, spent $12 and booked a dinner. Dayze turns that into a person, a calendar entry and a cost. You connect your accounts the way ChatGPT's connectors ask you to connect Gmail and contacts, and then ask your AI questions that draw on all of it.
"I just submitted my plug-in for ChatGPT and it got approved," he says, so you can search for Dayze in ChatGPT and install it now. A standalone app went in for review the day before.
The questions from the floor were about where all that data goes. For now, Luke says, it sits on his server, encrypted with a key for each user. He says only the person who logs in can read it. A builder asks, "You're sending it to a server, right?" Luke: "I mean, that is for now." The plan is an app that keeps the data on your own computer. The builder: "But still, like, I think it's risky to send it to someone else's computer." Luke: "True, it can be risky if you get hacked." A second builder asks when the data is encrypted and what happens when ChatGPT is processing their email. Anyone looking at the database without your key would see a jumble, Luke says. But through ChatGPT, you would see it. The exchange didn't get as far as when the email is decrypted, or what exactly ChatGPT receives.
Key takeaway: Dayze turns the accounts and notes you connect into one live context so ChatGPT (or Claude/Cursor) can answer without you re-explaining your life.
4. Pizza, Shy Window
Shy Window is a small Mac app that keeps a mini browser window floating on top of your work and slides it out of the way when your cursor goes near it. Pizza, a freelance designer who builds apps "mainly focusing on solving my own problems," talks about jumping between many chats with Command-Tab. Shy Window is Pizza's answer to that. "Actually the name should be Shy Browser, because it's a mini browser to help you more productive."
Pizza floats a reference website over the main window and turns on shy mode. Moving the cursor toward the corner where the window sits pushes it aside, so you can keep working underneath. The same works with a YouTube tutorial in picture-in-picture. The video keeps playing in the corner and moves away when the cursor needs that space. The last mode hides the video completely. "You want to see Tiktok but you don't want to let your boss see you," Pizza says. The video disappears, you look busy in ChatGPT, and when you hover over that area again it carries on.
A builder asks whether it installs on a work laptop. "Yes, can, it's a Mac app," Pizza says. There's no Windows version yet.
Key takeaway: Shy Window keeps a reference page or video floating on your screen and slides out of the way when your cursor needs that corner, so the reference stays visible while you work in the main window.
Applications
5. Anh Nguyen, a WhatsApp agent for Horizon Wellness Club
Anh built a WhatsApp agent that sits in a nutrition coach's group chat with clients, turns their meal photos and messages into food logs, and gives the coach a dashboard and weekly reports. The client is Horizon Wellness Club, which runs a physio practice and a gym and offers nutrition coaching. Anh started it a month or two ago, after attending Forward Deployment Academy here at SQ Collective, and calls it "probably 95% mostly working."
He first plays a video he made with Opus that walks through the pipeline. The agent joins the existing group and stays quiet. NanoClaw captures each message as it arrives and keeps the original. A later stage sorts the fragments into what belongs together, what's just conversation, and what's still unclear. The next reconstructs each client's whole day of meals, and code checks that the AI-generated meal records are properly linked and safe to save. Then he opens the dashboard itself. Anything the system wasn't sure about is clearly labeled for the coach to check. There are graphs of daily energy logs, trends over time, and weekly reports. The coach adds notes for the week against each client's goal, those notes go into the prompt that writes the report, and the coach can edit the report before producing the final PDF.
The design choice is keeping the WhatsApp group. Before this, clients sent meal photos to the group, and the coach downloaded every image by hand, organized them, and fed them to ChatGPT. There are plenty of AI food-logging apps, but Anh's read from this client is that "people don't actually like using bots or they don't feel accountable to bots." His take is that the coach is why people keep logging, so he keeps the agent working behind the group instead of replacing it. He added the sorting stage because the models kept making mistakes following the conversation, and the edit step because models "can't generate human reports consistently and accurately."
Anh said: "use AI in a way that still fits human behavior rather than just trying to kind of force fit AI into everything."
A builder asks how it accounts for body fat or BMI. Anh opens the add-client form to show it. The coach enters each client's relevant details and a clearly stated goal at onboarding, and those go into the model's context. For a lot of people, he points out, the goal is gaining weight or muscle, not losing it. Anh is a marketing consultant by trade, and when I ask how this ties back to his branding work, he says, "Actually I feel like it has nothing to do with what I did before." But this feels particularly empowering for him as a creative: he used to solve problems through branding. Now he can also solve them through software.
Key takeaway: Anh kept his client's clients in the WhatsApp group they already used and automated the downloading and sorting behind it, with anything unclear flagged for the coach to check.
6. Lucius, Pack These For
Pack These For is a packing-list site for hikers. Each guide covers one mountain: what to expect there, and a checklist of what to bring for that trip and season. Lucius is building it as the start of an outdoor safety platform for hikers.
The guide on screen is for the hike near KL (between KL and Genting) he planned to set out for at 10pm that night, to meet operators and share the idea on the ground. The guide opens with the mountain: 1,462 meters, eight to ten hours of hiking and 19.6 kilometers on the only trail you can take, in what he says is the wet season. Then comes the list, organized by what's provided, baggage guidelines, clothing, gear and equipment, personal items and first aid, and optional items for comfort, each with a checkbox to tick as you pack.
"When you're climbing a mountain, you never want to pack for the wrong season," he says. You can underpack on quantity, but you can never pack the wrong things for the season.
The idea came from his mom. A few days earlier, the two of them summited Mount Pulag, at 2,928 meters, her first mountain hike. She had no idea what to bring, and he wasn't sure either, since he'd never climbed it. The packing lists he found online are written as blog posts. "They expect you to read, but let's be honest, no one wants to read." Worse, "they never tell you how much you need to bring and why you need to bring that." That's what he wants his guide to say. Young hikers like him, he says, "tend to underestimate nature." Why bring a jacket? "Like, I can tahan."
From the floor, a builder who isn't a hiker says he'd use it for travel and asks for images instead of a text list. I add that I'd need a buying list, since I have no idea where the camping shop is. Lucius says hovering over an item could eventually lead to a buying guide and a way to make money, but "for now, I don't want to overthink it." The feedback he wants most is from tour operators, because they bring people up mountains and know what's essential.
Key takeaway: Lucius wants his list to say how much to bring and why, and he's taking it to tour operators to check what belongs on it before adding extras like shopping.
7. Deepan Krishnaa, PersonaFix
PersonaFix lets a marketing team create a reusable AI persona, based on a real person or made up from scratch, and then generate photos and videos of that same face in new scenes. I introduced it as deepfake territory. Deepan Krishnaa is its sole founder, and he's also a master's student at NUS.
He plays three examples. The first recreates a main character from HBO's Rome, made by uploading a couple of clips and asking for him in a toga. The second is purely synthetic, built from a one-line description of the kind of person he wanted. The third is a test for someone building an e-commerce store for activewear: his persona in a Lululemon yoga outfit, coming home, taking off a jacket and sitting down. It came from one prompt and two small refinements. Once a persona exists, a team can reuse it across video and other marketing material, so "they don't have to tackle with the prompts" each time.
"The images need a bit more work, in my opinion, because you can still tell this is AI." A builder asks whether he watermarks the images. "I haven't released this to the public yet, but I'll be adding a C2PA watermark," he says, so the output can be traced. Access will also be limited to enterprise, meaning people with registered companies, and he doesn't want random people using it.
Key takeaway: PersonaFix is built so a marketing team can reuse one face (real or synthetic) across new photos and videos without re-prompting the look each time.
SQ Collective hosts Coworking Fridays for founders, operators, and AI builders working through real product questions in Singapore.
Join an upcoming Coworking Friday, Technical Homebrew or a Masterclass: https://lu.ma/ai-labs
Explore SQ Collective: https://www.sq-collective.com
Comments