Founder @palettefm. indie ML researcher. former resident at @googlearts.
Porto, Portugal
Joined July 2010
- Tweets2.5K
- Following1.8K
- Followers13K
- Likes11.7K
Some of my early work in AI colorization is now on display in Vienna at Belvedere and outside the Medical University of Vienna.
tldr, Gustav Klimt's three Faculty paintings were lost during the second world war, and I collaborated with Google and Belvedere to colorize the remaining black and white photos of the artworks using AI.
More info:
- The Belvedere Exhibition (until September 2025): belvedere.at/en/gustav-klimt…
- Photos of the exhibition: flickr.com/photos/belvederew… © Photo: Ouriel Morgensztern Belvedere, Vienna
- Photos from Medical University of Vienna: flickr.com/photos/meduniwien… © MedUni Wien/ APA Fotoservice/ Hörmandinger/ Schedl
- Article about the project: artsandculture.google.com/st…
- Video summary: youtube.com/watch?v=1xYpIM_B…
Jut cut my AWS bill with 82%, and sped up my ML inference from 8 sec to 0.02 sec.
My tips:
- Use TorchServe with TensorRT, batch size 4
- Deploy with FastAPI, Gunicorn (w/ Uvicorn worker), and Nginx
- AWS auto-scaling cluster (NVIDIA A10G), ECS, LB w/ WAF
- Reserved instances
What does your ideal studio look like?
I’m aiming for an industrial yet homie vibe with a view of mountains and the ocean. And a few spare places for visiting creatives and hacker neighbours.
Emil Wallner retweeted
#AIart has never been about text2img, it's just the latest model to join a space with a rich and long history. Perhaps one of the best resources on this is @EmilWallner's ML Art project: mlart.co/, where you can find a plethora of artists and technique combinations
bumping into a twitter mutual deep inside a git issue is like crossing into an old friend on the street
every tech documentation should have a button to copy the entire doc so you can paste it into a chat engine and ask questions and generate templates
feels like siri should pivot to a voice-based app store. unbundling the voice parser, LLM backends, domain logic, and APIs. apple does hardware and systems really well but seldom end-user experiences.
if someone works in AI and sends you an email, it’s often more honest and concise to reply with a prompt: “a wholesome and personal decline to [email content]” via text-davinci-003
yesterday was my last day at google. i’ve had a wonderful time and made so many new friends 🫶🙏
today is my first day as a full-time indie ML researcher and indie hacker
student: write a 3-page review of the lord of the flies in the style of a B+ senior high school student
teacher: analyze the below review and extract the key points, write a paragraph of feedback on its format and originality, and score it from Fail to A+
I'm thinking of using this for my production ML serving. Any feedback or suggestions?
- AWS Loadbalancer + HTTPS
- Sagemaker Endpoint
- Uvicorn
- FastAPI
- TritonClient Py
- Triton Server (on Nvidia A10G)
- Nvidia Dali (pre/post-process)
- TensorRT via ONNX
these are the best articles i’ve found so far:
- towardsdatascience.com/huggi…
- towardsdatascience.com/servi… (@tanpengshi)
- francescopochetti.com/blurry… (@Fra_Pochetti)
- photoroom.com/tech/stable-di…
Extra plus for @pommedeterre33 ‘s git repo 💯