iiinova — where your digital self keeps evolving

Ramith Hettiarachchi @ramith.fyi · 25/07/2026

interesting looking one alphafold.ebi.ac.uk/entry/AF-A0A...

Ramith Hettiarachchi @ramith.fyi · 19/07/2026

been building an architecture explainer/explorer tool where you can go in varying depth of abstraction to understand: - started with the genie3 (in progress), then go towards AF2 Underneath, it's a compiler for turning model architecture descriptions into interactive diagrams scope.new.ramith.fyi

220

Ramith Hettiarachchi @ramith.fyi · 11/07/2026

Glad to have worked with this amazing team in organizing the GenBio workshop! ❤️ (There’s lot more who helped behind the scenes and couldn’t join in this picture) GenBio has been a great venue to bring researchers and foster very interesting convos.. wish I had more time to talk to even more ppl 😆

130

Ramith Hettiarachchi @ramith.fyi · 07/07/2026

Registration line 😅 even after two days of registration.. ml conferences maybe reaching a breaking point 😆 i guess it makes more sense to have more decentralized events

010

Ramith Hettiarachchi @ramith.fyi · 05/07/2026

First impression of SK: everything has a touch of cuteness ^_^

110

Ramith Hettiarachchi @ramith.fyi · 20/05/2026

*noise aware training* could make devices more resilient to fabrication imperfections in future (as shown in a different paper [1]) interesting how any kind of ML model learns other redundant shortcuts when it is constrained to work with noise [1] ieeexplore.ieee.org/abstract/doc...

060

Ramith Hettiarachchi @ramith.fyi · 08/04/2026

noticing that backbone design models trained on AFDB 588K, when sampled through ODE results in poor designability (~25%). But through a proper noise schedule in SDE it goes to 90% (as reported in lit.) this felt counterintuitive, wonder if the reason is, AFDB being just 42% designable to begin with

240

Ramith Hettiarachchi @ramith.fyi · 15/03/2026

training a flow matching model for protein backbone design (N, Ca, C) intra residue geometry is somewhat easy to learn, but inter-residue (peptide bond length, Ca-C-N and C-N-Ca angles) doesn't consistently match ideal stereochemical params [1] [1] www.ebi.ac.uk/thornton-srv...

011

Ramith Hettiarachchi @ramith.fyi · 10/03/2026

going to add this to CLAUDE.md Hard bugs: When stuck, write a detailed prompt and delegate to an agent rather than debugging directly. Crafting the prompt forces you to organize the evidence, which often matters more than who solves it.

020

Ramith Hettiarachchi @ramith.fyi · 18/02/2026

starting to love the @marimo.io notebook's reactive cell execution. all the nodes in the DAG are on the same page!

140

Ramith Hettiarachchi @ramith.fyi · 11/02/2026

ok nvm, i'm so stupid.. sometime back I restricted sshd to use only core 0 🤦‍♂️🤦‍♂️🤦‍♂️🤦‍♂️🤦‍♂️🤦‍♂️🤦‍♂️🤦‍♂️ bsky.app/profile/rami...

000

Ramith Hettiarachchi @ramith.fyi · 11/02/2026

this is kinda stupid.. need to figure out why numpy codes only restrict to one core python -c "import numpy; import os; print(f'Allowed Cores: {len(os.sched_getaffinity(0))}')"

100

Ramith Hettiarachchi @ramith.fyi · 03/02/2026

Agree.. I tested a notebook with some tests at the bottom. Was able to run that file with pytest as well, which is helpful Also i noticed that the support for pandas dataframes is pretty nice.. (to make quick charts etc) -- not sure if jupyter allowed that

100

Ramith Hettiarachchi @ramith.fyi · 03/02/2026

@marimo.io's hover tooltips showing function signatures & docs is nice

120

Ramith Hettiarachchi @ramith.fyi · 26/01/2026

Yeah, atm typst doesn't have a markdown export.. it's designed to replace LaTeX..

110

Ramith Hettiarachchi @ramith.fyi · 26/01/2026

building a website through @typst.app is kinda straightforward.. was trying out this template: github.com/vsheg/tufted, forum.typst.app/t/tufted-web... for eg: Below you can see an example post from their template: new.ramith.fyi/posts/2025-1...

Screenshot of the Typst web application interface showing a split-screen view of a project titled "typst-website." The left panel displays a file directory with a "posts" folder expanded, revealing various dated entries. The center panel contains a code editor with Typst markup for a document titled "The Normal Distribution," including text, mathematical syntax for the probability density function, and code for generating a diagram. The right panel shows the live preview of the rendered document, featuring formatted text, the mathematical formula clearly typeset, and a blue histogram chart illustrating the distribution. The path bar at the top indicates the file location: Ramith > typst-website > content > posts > 2025-10-30-normal-distribution > index.typ

3383

Ramith Hettiarachchi @ramith.fyi · 11/12/2025

was hoping to see some reasoning traces on the math countdown task.. instead, diffusion LLM, started saying "exactly" too many times 🤨 Maybe its "reasoning" is during denoising (remasking + unveiling), rather than the autoregressive case where it explicitly says ("oops i got that wrong")

010

Ramith Hettiarachchi @ramith.fyi · 05/12/2025

This is the question in the validation set

010

Ramith Hettiarachchi @ramith.fyi · 05/12/2025

Trying to train a diffusion LLM with reasoning on a math task (without SFT), and this is what it spits out 😒😑

120

Ramith Hettiarachchi @ramith.fyi · 14/11/2025

Installing alphafold3 on the DGX Spark was pretty easy, but I forgot that JAX preallocates gpu memory* (XLA_CLIENT_MEM_FRACTION=0.95) Forgot to change those settings, and the spark froze without me being able to ssh 😆 *In the spark, it’s unified mem Gotto test again either reduced reallocation

100

Ramith Hettiarachchi @ramith.fyi · 05/11/2025

An application asked for the lifetime number of lines coded.. with gemini's help wrote a code to count net lines (.cpp or .py) an author contributed to a repo.. Thinking of polishing it up further: current version: gist.github.com/ramithuh/aa6...

110

Ramith Hettiarachchi @ramith.fyi · 02/11/2025

dgx spark training speed impression while training a diffusion language model on Shakespeare: DGX Spark: ~1.3 it/s L40: ~3.5 it/s (same batch size used here)

010

Ramith Hettiarachchi @ramith.fyi · 22/10/2025

when you cant miss out a dinner w friends, yet as a TA you have to answer some questions students have* 🥹😅 *for a HW that's due midnight

110

Ramith Hettiarachchi @ramith.fyi · 21/10/2025

one of the fondest memories I have during high school time is intel shipping me an edison development board.. back in 2014, i emailed a PM @ Intel with a suggestion, she kindly acknowleged it.. and the next day she asked for my address so that she could send me an intel galileo board.. (1/4)

A collage of screenshots showing emails from 2014-2015 and a LinkedIn message from 2022, documenting an Intel employee sending the user a 'Galileo board' in high school and the user following up years later to express his gratitude and its impact on his career

110

Ramith Hettiarachchi @ramith.fyi · 19/10/2025

Sometimes I cook so that I can listen to a podcast

000

Ramith Hettiarachchi @ramith.fyi · 14/10/2025

the package ydata-profiling is pretty nice github.com/ydataai/ydat... was doing some data analysis in the plinder dataset, generating a bunch of high level info with a single command

example output report from ydata-profiling.

010

Ramith Hettiarachchi @ramith.fyi · 30/09/2025

view from Eagle rock trailhead was breathtaking ❤️

020

Ramith Hettiarachchi @ramith.fyi · 24/09/2025

My takeaway: Their Approach == ESM2 (3B) + SimpleTransformer w. Flow Matching (X B) When X=100M, it falls way short of ESMFold performance (10-15%) When X=3B, it closes the gap ---> Still, AF2+MSA (95M) is far superior But it's nice to know how far simple transformer+flow matching can take us

120

Ramith Hettiarachchi @ramith.fyi · 30/08/2025

Good to be back in Pittsburgh ❤️

010

Ramith Hettiarachchi @ramith.fyi · 18/08/2025

mol-view-spec’s Jupyter and streamlit integrations are kinda cool github.com/molstar/mol-...

010

Ramith Hettiarachchi @ramith.fyi · 13/07/2025

haha.. thats pretty cool! 😆

1140

Ramith Hettiarachchi @ramith.fyi · 26/06/2025

kinda cool that the url has all the code, and nothing is stored on the server.. can even put in a git commit message (to explain something significant underlying the commit) 🤔 (although i'd be wasting 1KB for a commit message)

000

Ramith Hettiarachchi @ramith.fyi · 13/06/2025

It's nice to add a keybinding in Jupyter Lab for things like "restart kernel and run all"

000

Ramith Hettiarachchi @ramith.fyi · 05/05/2025

damn JVKE has written songs on how everything feels like 😆❤️

Screenshot of a YouTube results page showing six videos by artist JVKE in a “this is what … feels like” series.

100

Ramith Hettiarachchi @ramith.fyi · 26/04/2025

pymol looks cool with `bg_color [0.0, 0.0, 0.502]` sort of old computer interface background color scheme

010

Ramith Hettiarachchi @ramith.fyi · 26/04/2025

Const pH MD going weee 🏃‍♂️

110

Ramith Hettiarachchi @ramith.fyi · 23/04/2025

ok upgrading deepspeed worked.. special thanks to perplexity and gemini for validating my choices before running them

000

Ramith Hettiarachchi @ramith.fyi · 12/04/2025

Had some fun for homework, running a pipeline composed of virtual screening (pharmit) + pocket conditioned ligand design (DrugFlow --> this was easy to setup/run 💯)

The figure visualizes a candidate molecule, DeNovo β (green sticks), docked into the binding pocket of the energy-minimized NRAS Q61R-GDP structure.

220

Ramith Hettiarachchi @ramith.fyi · 21/03/2025

Interesting how far you can go when you just focus on improving one step at a time 😅 A reminder for me to apply this to other aspects of life 🤞

110

Ramith Hettiarachchi @ramith.fyi · 15/03/2025

Slight improvement from last week 💆🏻‍♂️ 🧗

000

Ramith Hettiarachchi @ramith.fyi · 14/03/2025

@warp.dev's terminal is pretty helpful..

A picture depicting warp.dev terminal which has some features to ask natural language queries in a terminal

110

Ramith Hettiarachchi @ramith.fyi · 08/03/2025

New bouldering challenge for the rest of the month 🥹😅 #bouldering

130

Ramith Hettiarachchi @ramith.fyi · 01/03/2025

Hoping to get this right the next time -_- #bouldering

250

Ramith Hettiarachchi @ramith.fyi · 28/02/2025

looking forward to a day where I can make an actual figure for a paper 😅🤞 (this was for an assignment)

Figure 1: (A) Structural overview of NRAS in complex with GDP (inactive conformation), highlighting key functional elements: the P-loop (purple), Switch I (green), and Switch II (cyan). (B) Superimposed WT and Q61R NRAS structures color-coded by root-mean-square fluctuation (RMSF), illustrating regions of differing flexibility. Note the shift in RMSF (red scale for WT, blue scale for Q61R) around residues shown. There seems to be a drastic difference at TYR-33, where in the mutant the flexibility is low (cyan, instead of purple). (C) Distance fluctuations (D1 and D2) along the molecular dynamics trajectory of WT and Q61R is shown. The solid line depicts the moving average (window size = 200), for ease of visibility. D1 is the distance between the P-loop and Switch I and D2 is between the P-loop and Switch II. It can be seen that for both D1 and D2, the distribution is spread out while having a bi-modal distribution, meanwhile the mutant has a narrower single mode distribution. By looking at the distribution of Q61R, it appears that it is “stuck” in a conformation where the distance between p-loop to SW-1 is low. This observation aligns with Q61R's known role as an activating mutation that locks NRAS in its GTP-bound active state.

000

Ramith Hettiarachchi @ramith.fyi · 24/02/2025

when I first made a Visual Basic window application as a kid, I felt like a "Software Engineer" Now having worked with pymol/MDAnalysis a bit, I feel like a "Structural Biologist" 😆🤦‍♂️ same vibes, different times 😅

A 3D structural overlay of the NRAS protein is presented, with the wild-type (WT) structure in yellow and the Q61R mutant in white. The visualization highlights regions of high RMSF values during molecular dynamics (MD) simulations. These high-fluctuating regions are marked in red for the WT and cyan for the Q61R mutant.

110

Ramith Hettiarachchi @ramith.fyi · 13/02/2025

this is fun to watch.. 😆 learning flow matching by trying to overfit to one data point 😅

120