The idea for Commissioned started taking shape during a hackathon for Software Engineers that me and Gabe participated in. And like many of the best ideas that come out of hackathons, it wasn’t the project we submitted but a product of a problem we noticed. Out of all the dozens of impressive projects at that AI themed hackathon, we were the only ones using fine-tuning.
That was quite surprising to us because we knew how valuable it can be. Even in the hackathon, our project’s eligibility model had gotten a 35% boost in accuracy from fine-tuning. But something else also became obvious: fine-tuning is intimidating even for seasoned developers.
When we looked closer, the reasons became clear. People experimenting with fine-tuning kept running into daunting data formatting requirements, brittle pipelines, and painful infrastructure setup. These time-consuming, low-level tasks have shaped the perception of fine-tuning as something slow, expensive, and reserved for large companies with dedicated ML teams.
We built Commissioned to change that. We want marketers to fine-tune models that write in their voice without having to know what a JSONL is. We want ops teams to get better labeling and classification models by simply uploading a CSV. And we want developers to focus on building differentiated products instead of rewriting the same integration logic for every provider.
Our goal is simple: make fine-tuning accessible enough that anyone can experience the value it can bring in minutes not weeks.
So it's always a little annoying to find pictures on your device especially if you have many. Google Photos does an okay job if you label everybody but it still wouldn't get something like as specific as "Grandma in her iconic red sweater and round glasses". So, I made a search tool that does it without tags or labels. This is a super early version, so just trying to get feedback. Couple notes: it's limited to 50 pictures and will always return the 5 "closest" images to whatever you type.
That was quite surprising to us because we knew how valuable it can be. Even in the hackathon, our project’s eligibility model had gotten a 35% boost in accuracy from fine-tuning. But something else also became obvious: fine-tuning is intimidating even for seasoned developers.
When we looked closer, the reasons became clear. People experimenting with fine-tuning kept running into daunting data formatting requirements, brittle pipelines, and painful infrastructure setup. These time-consuming, low-level tasks have shaped the perception of fine-tuning as something slow, expensive, and reserved for large companies with dedicated ML teams.
We built Commissioned to change that. We want marketers to fine-tune models that write in their voice without having to know what a JSONL is. We want ops teams to get better labeling and classification models by simply uploading a CSV. And we want developers to focus on building differentiated products instead of rewriting the same integration logic for every provider.
Our goal is simple: make fine-tuning accessible enough that anyone can experience the value it can bring in minutes not weeks.