You are browsing as a guest. Sign up (or log in) to start making projects!

advay_awesomer

@advay_awesomer

Joined June 1st, 2026

  • 3Devlogs
  • 4Projects
  • 3Ships
  • 30Votes
Open comments for this post

1h 2m 41s logged

I designed and built pulse, a calm vanilla-HTML/CSS/JS new-tab page (clock-synced ECG “heartbeat,” search, speed-dial, Open-Meteo weather, focus timer with streaks, Cmd-K palette, theming) on a custom design system.

0
Original post
@advay_awesomer

I designed and built pulse, a calm vanilla-HTML/CSS/JS new-tab page (clock-synced ECG “heartbeat,” search, speed-dial, Open-Meteo weather, focus timer with streaks, Cmd-K palette, theming) on a custom design system.

Replies

Loading replies…

0
3
Ship

Made chesscnn — a chess engine that picks its moves from a convolutional neural network instead of hand-written rules, AlphaZero-style. The net (PyTorch, a residual tower with two heads) looks at the board as a stack of 19 image-like planes and outputs both a move-probability map and a single "who's winning" score. A PUCT Monte-Carlo Tree Search then uses those to actually play. There's a CLI and a Streamlit web app where you can play it and watch its eval bar, its predicted line, and a heatmap of where it wants to move.
Challenging: encoding chess moves for the network was the rabbit hole — every possible move has to map to one of 4,672 output slots (the 8×8×73 scheme), and promotions/castling/en-passant all have edge cases. I round-tripped 219k moves to prove the mapping never collides. The other hard part was getting it to play well without a giant training run — a freshly-initialized net plays nonsense, so I blend the network's evaluation with a classical hand-crafted one and let the search do the heavy lifting.
Proud of: it plays real openings out of the box (it'll walk into a Ruy Lopez or a Pirc), it found a tactic in testing (won a hanging queen), and it went 0 losses across 14 games vs a random-mover. ~9,000 lines of Python with 53 passing tests, and a full training stack (self-play → replay buffer → trainer → arena/Elo) if you want to make it stronger.

To test it: clone github.com/advaycode/CNN1, pip install -r requirements-app.txt, then streamlit run chesscnn/app.py and play it in your browser.

  • 1 devlog
  • 55h
Try project → See source code →
Ship Changes requested

Made GNN-PCNA: a graph neural network that scores which residues on PCNA — a notoriously hard cancer drug target — are worth
investigating as hidden ("cryptic") pockets. It reads a protein as two graphs at once (3D contacts + backbone chain) and fuses them
with a learned attention gate — a GATv2 design the standard baseline (PocketMiner) doesn't use.

Hardest part wasn't the model — it was catching homology leakage in my own data. My first benchmark looked great until I ran MMseqs2
and found near-identical proteins (99% identity!) sitting across the train/test split. So I deleted my own headline numbers and
rebuilt everything on a clean 30%-homology split. The honest score came out lower — and that's the whole point.

Proud of: it's engineered to not overclaim. Every output is framed as "hypothesis generation," an automated verifier catches my own
mistakes, and an ablation proves how much comes from sequence vs. structure.

To test it: open the demo, pick a PCNA structure (try 8GLA or 1W60), and watch it score every residue in ~2s with a 3D viewer +
heatmap. Heads-up: the live demo runs the lightweight model; the headline AUPRC/AUROC come from a heavier ESM2 model (too big to host
free). Full story in the repo's DEVLOG.

  • 1 devlog
  • 13h
Try project → See source code →
Open comments for this post

13h 23m 59s logged

first making a memory base for the gnn

0
Original post
@advay_awesomer

first making a memory base for the gnn

Replies

Loading replies…

0
3

Followers

Loading…