Ananya Pathak
@AnanyaPath45073
Followers
38
Following
420
Media
34
Statuses
197
ML Engineer | Bio-medical Engineer | Building https://t.co/xVsiMVPQnH | Experimenting with ML @ https://t.co/g9UlKwt1bE
India
Joined March 2023
All my sites are up and healthy. Yay!
0
0
1
Dashboard coming in great. Changed to logo. added functionality to change appearance SMTP server to send mail for forget password functionality. How does the logo look like though?
0
0
0
Point number 3 and 4 are mandatory for vibe coding. The best thing is you can vibe code the ci/cd testing as well
1
0
3
These are the softwares I self host on my VPS: 1) Coolify - open source vercel by @heyandras 2) Plausible - open source alternative to GA 3) My personal website 4) My SaaS - includes postgres, adminer, fastapi backend, workers, redis contd.
2
0
3
TIL The new Image gen tools like Nano-banana are really good with making patent drawings especially line drawings and method flow charts
0
0
0
PatViz now supports patent standardization, cleaning and merging results from USPTO and Espacenet. Quit paying 1000's of USD for expensive patent software. Try PatViz for free
0
0
2
Cold emailing Back in 2016-2017 started emailing profs in Europe for internships. Luckily I got a paid position for 4 months in France. Learned more doing this than the internship itself 🤣
1
0
5
87% bounce rate for the last week for PatViz( https://t.co/Tw06apxDaf) How do I improve this ?
0
0
2
Changed the patent mapping dashboard on the patviz app 1) sidebar instead of a header 2) interactive form based on user input 3) Support for time series charts Working on customizing charts and sharing ! site: https://t.co/5E9y2a1T29
0
0
1
Lovable + cloudflare pages is almost a super power. - Lovable - build and push to a github repo - Cloudflare pages - automatically picks the main (customizable) branch and re-deploys it. logs , uploads and cloud functions can be configured from the dashboard
0
0
0
Been playing with time series graphs now. Plotly does offer a great visual UI ! https://t.co/5E9y2a2qRH
wasted 2 days on troubleshooting this
0
0
1
some great results on swapping self attention with fourier transform: 1) ~33% and ~9% faster inference on CPU and GPU 2) achieves ~96% accuracy of the original model (bert-tiny) 3) ~43% and ~23% faster training on CPU and GPU
fnet bert tiny vs bert tiny (vanilla) on imdb dataset. 25 k training data pts, 3 epochs each. fnet has faster training and inference time. PS: trainable params vanilla bert tiny - 4,416,698 trainable params fnet bert tiny - 4,284,090
1
0
2
Def the best piece of tooling if you want to vertically scale your web app
0
0
0
this is actually decent for niche use cases on edge devices. Will experiment with 1) training for more epoch's (my guess is that the gap between accuracies will decrease ) 2) on more complex datasets
0
0
0
fnet bert tiny vs bert tiny (vanilla) on imdb dataset. 25 k training data pts, 3 epochs each. fnet has faster training and inference time. PS: trainable params vanilla bert tiny - 4,416,698 trainable params fnet bert tiny - 4,284,090
Doing a fun experiment: Swap the self attention layer of a bert based encoder with fourier transform. Finetune on a small classification dataset and observe the trade offs bw accuracy and inference speed. Something like Fnet
1
0
0