Mike Xia
mike.xia [⍶t] mit.edu
🐦 @0x796F
▲ new opportunities
looking at generative ai or ai enterprise tools, but also open to fast growing products that leverage my prior experiences
▲ investing
ai saas, developing markets, and some hardware products with high attach rate mrr, 5-10k check on a safe
▲ advise / consult
most helpful on early stage eng, go to market, anything hardware / design, but there’s a 20/80 for everything else
search 16TB of images in 10ms
built unprompt.ai to quickly find prompts for any style of ai generated images.
scraped and indexed 16TB of image data using the clip encoder and added some filters for aesthetics and nsfw.
this delivered a simple yet powerful search interface on ai art. kinda viral @ 60k mau
2023 design awards for lumina
we decided to apply 2 years later (mostly because in 2021 the co was too poor to pay the fees)
camera’s are complicated, way more than i imagined. Lumina won both the iF and reddot design awards, way to go team!
$7 per day = 3000 cold emails
after years of being a technical founder, finally put together a simple sales stack. its very 20/80, designed for 1 or 2 founder led sales via cold email outbound.
you can mix and match scalable sources such as linkedin and crunchbase, making this useful for more than just b2b sales (ie. fundraising, recruiting, networking)
heres an ai generated image for lulz
building a DSLR camera
lumina is a DTC 4k webcam that uses AI software to replicate how a DSLR looks. conf call apps only send 25fps 720p, so for 99% of calls a DLSR is overkill.
challenge is getting the AI to run on fragmented consumer devices, in hind sight no wonder vc’s didn’t want to fund this.
the company now has great ebitda, and a strong team of 20+ pushing the needle
fpga miner design
for 2 years i got really deep into fpga mining of pow altcoins.
we shipped recycled pcbs from the US to shenzhen, useful chips were removed with a heat gun, cleaned, reballed, tested for functionality, and re-assembled into altcoin mining machines.
targeted EOL 28nm chips, 95% util, folded cores, single wire comms, ram as LUTs, 600-700 mhz without errors, graphene coated heatsinks.
god it was insane.
ai for k12 teachers
way too early for the ai party
was a first time founder who used RNNs back in 2015 for offline handwriting recognition and tried to sell this technology to teachers and schools.
technology was fun to build, but as a first-time founder made a ton of mistakes. geoff and tim from yc took a bet on us thanks guys!
famo.us
demo used css 3d matrix transforms to animate dom elements using GPU acceleration. right side you can see an image animation built entirely in DOM lol.
no business value but wow, full demos will be uploaded soon. just need to find time …
mit lab
senseable.mit.edu
built a cocktail making robot, went to milan design week, lotta drink.
did field research on cape town’s jitney taxi system, 20 people a van, music blasting, dirt roads, whew. wine was amazing
other stuff
Mike Xia 