Unwavering for a Decade

After a decade-long hiatus from blogging, I’ve decided it’s time to share my story of perseverance and determination in building a venture. A recent morning’s reflection sparked this idea to write again - while I’m not sure if I’m ready to share my story, I know I should start writing it down.

The day zero

How did this start? I remember facing a career choice in 2014: continue as a post-doc at UCL pursuing academic excellence or become a founder chasing my own vision.

I chose the latter, because I love to create, build, and share real things.

From 2014 to 2020, I co-founded Umbo Computer Vision. It was a journey of building a company from scratch, sharing my technical expertise as CTO, but ultimately failing to convince myself to continue as a co-founder. Deep Learning based Computer Vision algorithms were a new frontier at that time (i.e., Convolutional Neural Networks, a.k.a. ConvNets, such as AlexNet, VGG, ResNet, YOLO, etc.).

I was fortunate to be part of a startup building the first AI product, a real-time object detection system for autonomous video security. We shipped AI products running on both cloud GPU and edge AI ASIC platforms. I gained experience in building large-scale AI systems and production-level AI products. It was also a $9.2M-raised lesson in learning the hard way that solid technology doesn’t necessarily speak for itself. Go-to-market is the key. I left the founding team due to conflicts over long-term company vision.

In June 2020, I started Instill AI with my UCL lab colleague, Xiaofei. We had worked together in the research lab for years, and she had joined Umbo after completing her PhD. We shared the same vision of how AI should be implemented and adopted as a software infrastructure. The initial article addresses Instill AI’s vision in more detail.

It has been 4 years since founding Instill AI. Our effort focuses on executing a firm belief in unstructured data ETL. Processing unstructured data should be as easy as processing structured data. Instill Core has been our open-source initiative to realize our vision of an ideal unstructured data ETL tool. We’ve been quite opinionated about its design and architecture. See this article for more details.

Looking ahead to the future

Since 2020, we’ve raised $4.2M from top-tier investors, and Instill Core (and its fully managed service Instill Cloud) has been accessed by more than 1,500 users.

In 2025, the AI and data market will be booming. Large language models (LLMs) and large multi-modal models (LMMs) are the new frontier of AI and have significantly changed software practices, especially in the unstructured data space. In my opinion, compared to the first wave of ConvNet-based Deep Learning, this is primarily due to:

• Pre-trained Transformer-based models being powerful enough to handle unstructured data in general cases
• Scaling laws having worked well in recent years
• Giants fighting to monopolize the market, accelerating the commoditization of foundation models and infrastructure (similar to what happened in the Cloud and Big Data era)

It’s a great time to be an AI and data founder. In 2025, we’ll see more success stories about AI assistants and agents in the application layer, as scaling laws reach their limits and retrieval augmented generation (RAG) becomes common practice for preventing LLM hallucinations. I’m excited to be part of the party.