Friends+Family DirectAI Monthly Update (February 2024)
Hey all!
Keeping you up to date with what’s new at DirectAI. To reiterate, we’re enabling anyone to build computer vision models that classify, detect, and track objects in images & videos, without training data. If you missed last month’s update, check it out here!
Last Month’s Rundown
Lost a contract we’ve been working on since December. More details in learnings.
Closed two small recurring revenue customers. We’re doing customized content moderation for a social media company. We’re also doing object detection for a home renovation company that ingests & transforms images of home exteriors.
Lots of outbound. Lots of prospective customer calls. We released native support for RTSP Streaming and improvements on our classifier product!
What We Learned
Selling to re-sellers is slow and signal is lossy.
We’ve chatted with a number of Video Management System and/or Camera companies that want to sell video analytics to their customers. They say they want to pay for DirectAI but want to show a video analytics demo to their customers beforehand. As we’ve come to find out, this is a kiss of death. This class of prospective clients does not have a willingness to pay. Our tech doesn’t solve their problems. They are neither the ultimate user nor the ultimate buyer. The camera company we’ve mentioned in the past few updates falls pretty cleanly into this definition. We lost a contract with them because of a miscommunication over their end customer’s actual needs.
People want models they can run on their machines. They want to configure models quickly, develop conviction on an implementation, then deploy it on edge-devices without frequent updates.
We talked to a number of camera companies that have custom computer vision problems. They’ve either been burned by prohibitively expensive cloud solutions (GPU costs for large foundation models add up when you’re analyzing hundreds of streams) or they only have access to edge-deployed compute (e.g. Jetson Nano, Ambarella, Coral). They go through the arduous process of training their own models or deploy tiny, pre-trained, open-source models for constrained tasks. We think we can meaningfully accelerate the engineers working on both tasks.
Sales calls are user interviews too.
It’s great to ask people what their problems are without the expectation that the conversation will lead to a commercial partnership. But you can get even clearer information about the magnitude of people’s problems when you ask them upfront whether they’re willing to pay to have them solved. We’ve gotten clearer signal from running an outbound sales campaign than from conducting an equivalent set of coffee-chat user interviews.
One prospective enterprise client offered to begin the due diligence process in hopes of acquiring us. This was initially quite flattering but we realized that, while selling your company makes sense above a certain price, 1) we didn't have strong conviction that selling DirectAI was the best path forward and 2) the acquisition process is as distracting, if not more so, than raising a new round of funding. Without strong negotiation leverage (I.e. term sheet or competing offers), we would expect a lowball offer.
Attaching some resources which we used to inform our decision not to begin that process: PG’s Don’t Talk to Corp Dev (and HN’s perspective), The Founder’s Guide to Selling Your Company, as well as this Blog Series on Getting Acquired.
What’s Up Next
We’re still moving a few camera pipelines through the pipeline but are shifting our focus to the tiny model hypothesis. Namely, we think we have enough product differentiation that it’s worth going back to robotics teams that we chatted with last spring. Much like last month, we’ll be doing a time-bound outbound sales motion.
Asks
We’ve requested this before, but we’d really appreciate introductions to robotics teams that are deploying vision models on tiny GPUs.
Kindly,
Ben
P.S. Isaac and I celebrated our birthdays in February!