The Demo Day S4

The room at SQ Collective is quiet, save for the hum of laptops and the occasional shuffle of chairs. At 1.30 the AI Labs came to life, and builders took turns stepping up to the screen to show what they have been hacking together over the past month.

At SQ Collective, we organize these monthly demos to encourage everyone to move from vibe to real feedback.

AI for Fashion, but one that actually understands you

Yu Qing steps up first. She is building an AI styling and shopping companion. She starts by pointing out how current platforms push inventory based on trends, not based on the person.

"We're not going to focus on the inventory of the brands," YQ says. "Because the brands want to push you with something because the trend tells them to do so."

Her app instead tunes into mood, life chapter, and body type to enable hyper-personalisation at scale. It handles the nuances of daily life.

For women shoppers, she explains, the AI considers the physical reality of the user, even to details like the dates of the lady’s menstrual cycle. It knows if a period is coming up in three days, anticipating the days you might feel bloated and want to avoid tight clothing or light colours. The closet planning changes entirely.

The goal is to shrink the morning decision cycle from 30 minutes to three minutes.

Someone from the audience chimes in to ask about a camera function for virtual try-ons.

Yu Ching acknowledges it is on the roadmap but grounds the project in immediate utility. "Right now, let's take care of the basic, what we feel on the daily basis, and then we'll move on to a long term," she says.

Democratising the Bloomberg Terminal

Next up is Ali. He used to work at Bloomberg and noticed a gap. Institutions have terminals with massive analytical power. Retail investors only have basic brokerage apps.

He boots up a localhost application. It is a portfolio rebuild that brings institutional tools to the retail level. You upload your portfolio, and it breaks down your positions against benchmarks alongside live macroeconomic indicators.

But the core feature of this iteration is what he calls the shock test.

You can write an event in natural language - like a blockade in the Taiwan Strait or a sudden political death - and the system runs shock tests against your live portfolio to see how it would hold up. It asks different AI council pods to evaluate if your current strategy is sound under those specific constraints.

A question comes from the floor. An attendee presses Ali on the baseline of these tests, asking if the LLM is just spewing from past wrong data since the market moves so fast.

Ali clarifies that the system pulls live prices down to the minute. It takes a precise snapshot of the market to run the simulation. Another builder in the room nods, noting that they are working on a patent that adds a dynamic layer to it. They agree to talk after the session.

Reasoning Over Execution

Arunak takes the next slot to show his agentic harness for quantitative Finance.

He has built an open-source architecture that separates reasoning from execution. The LLM only handles the reasoning layer. Underneath, a deterministic software layer with over a hundred tools executes the financial metrics.

He demonstrates a workflow that targets quantitative researchers. The system ingests a dense research paper: he uses a famous paper on Warren Buffet's alpha as an example, and the system automatically extracts the methodology. It then dynamically generates the strategy, writes the code, runs a backtest on cached stock data, and produces a replication report.

"LLM is sort of doing the inference and the response generation and deciding what to use," Arunak says. Everything else is deterministic code. This provides a strong audit trail for researchers, highlighting discrepancies and transaction costs without the black-box opacity of a standard chat interface.

Someone in the room asks about benchmarks. How does he prove the performance of the harness, and is there an industry-specific benchmark he uses.

Arunak admits that is still in development and he will likely have to build the benchmark for the finance vertical himself. He is confident about writing the evals for the AI outputs, but to make it really useful, it will also require feedback from domain experts themselves.

The attendee immediately offers to help him build those benchmarks, and they make plans to connect.

Finding the Signal in the Noise

Hansel is up next. He is not technical, but he built a go-to-market tool for a Series B quantum sensing startup. But his experience with venture building helped him to really zone into the business logic behind B2B outreach.

His client, a startup that produces high tech quantum sensors for the detection of minerals below ground, needed to sell to mining companies. The experts that understand and are willing to buy these technologies don’t appear on traditional social media.

However they are hidden in the thick pages of public disclosures of the Australian Stock Exchange. What’s not easy is that there are over a thousand miners listed there. Their profile and purchasing activities are buried in 60-page PDF announcements.

Hansel built a pipeline that processes 40,000 PDFs to find the top 25 target companies. It scores them based on highly specific signals, like whether they have partnered with a tech startup before or if they have previously attempted gravity-based exploration.

He pulls up a dashboard showing the top individuals to contact within those companies. In a high stakes business development activity like this:

"You have one shot and if you hallucinate, you're done," Hansel says.

To solve this, every signal in his system is tied to an evidence log with verbatim quotes from the original source documents. When the two-person business development team reaches out, they base their message on verifiable facts.

Takeaway for me: the competitive edge is not the code. It is deeply understanding the business logic. AI's value multiplies when you can clearly define what "good" looks like in your specific domain and then build a system to scale that exact expertise. The technology is just the lever for domain knowledge.

An audience member asks how portable this system is across other industries.

Hansel is straightforward. The architecture - the scoring, the decay rates, the tracking is portable. But the signals are entirely custom. That non-portability is exactly what makes the tool valuable. No one else has that specific edge.

A self-learning LinkedIn Content Engine

Anh Nyugen showcased a brilliant solution: a personalized AI content engine designed to cut through the noise. This innovative system, powered by an AI agent named "Claudia," drastically reduces content creation time from what took him five hours to just one, while maintaining his consistent, authentic voice.

Here's a sneak peak of his workflow: Claudia conducts weekly research based on predefined content pillars and populates a Notion content hub. Crucially, the AI is trained on Anh's past posts to mimic their unique writing style, ensuring content never sounds templated.

To add a personal touch, the system prompts Anh with specific questions, AKA his hot takes, allowing for a "brain dump" of individual insights before drafting.

While the AI generates the initial draft, Anh emphasizes the vital role of human oversight for final edits, ensuring every post resonates with genuine subjectivity. The copy that is eventually posted is fed back into the AI, checked for differences, and updated into the learning loop.

This blend of AI efficiency and human authenticity is a game-changer for LinkedIn engagement.

Streamlining On-Demand Manufacturing

In the world of on-demand manufacturing, a significant bottleneck often arises: the manual, time-consuming process engineers face when identifying, understanding, and validating parts.

Ezekiel introduced a digital inventory management system designed to tackle this very challenge, promising to mitigate supply chain disruptions.

It automates the entire workflow, from identification to ordering, in four key steps. It leverages advanced AI with semantic understanding to accurately identify components, distinguishing, for instance, between a "pump" and a "pump motor" from procurement data.

This intelligence allows the system to extract all relevant information and create a digital twin entry for a part in just ten minutes, a task that could otherwise take an engineer half a day.

Furthermore, it optimizes the ordering process by replacing specific product titles with more generic ones, simplifying fulfillment through decentralized suppliers. This innovative approach not only saves valuable engineering time but also enhances the efficiency and resilience of on-demand manufacturing.

Changing the Way We See

Sophie is a product manager who only recently started exploring interactive storytelling after joining the SQ community. She has always had an interest in scientific story telling, but the tools for creating these compelling stories were not easy to learn in the past.

She presents a visual story about John Snow. Not the television character, but the British physician who challenged the medical consensus during the 1854 cholera outbreak in Soho.

At the time, people believed cholera spread through bad smells. Snow mapped the deaths and found a cluster around a specific water pump.

Sophie scrolls through her project, the map dynamically drawing the data points.

"When did the way of seeing change," she asks the room. "Was that when he started mapping."

But the map did not change minds. The medical body rejected his findings. Meanwhile, other reformers continued building sewage networks based on the wrong theory of bad smells - which accidentally saved lives anyway. Snow's theory was only accepted 12 years later, long after he had died.

Sophie shares what this historical footnote means to her as a builder. We might be acting on the right theories and never see the validation, or we might be acting on the wrong theories and still doing good.

"I felt like I don't have to be held hostage by needing to know," she says. "And then maybe I will just do what I can do. And then continue."

The room digests that quietly. It is a good reminder for anyone building at the edge of a new technology. You build your map, you test your assumptions, and you put the work out there. The recognition might catch up later, or it might not. The only thing you control is the work itself. Keep building.

---

This is the kind of room where people who are quietly building the future come to show their work. It is a space to test ideas, get real feedback, and find collaborators who understand the mechanics of what you are trying to pull off.

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

Usually over laptops. Sometimes over pizza.

Join the next one.