Skip to content
View Feonixid's full-sized avatar
💭
Doing nothing about my sleep
💭
Doing nothing about my sleep

Block or report Feonixid

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Feonixid/README.md
Typing SVG

Status: doing nothing about my sleep.

The Balkans  ·  English · Albanian · Greek

Who am I?

I'm Feonixid (Laerti Shabani) — a student and an independent ML researcher from the Balkans. Everything I build is connected by one simple thread: AI. I want to understand how intelligence emerges from training and move the needle closer to AGI one small step at a time.

My entire lab is two Kaggle T4s and a bit of obsessiveness.

Research

Every project ships with code, raw per-seed results, and a single canonical reproduction script. Failed runs and negative results are reported, not hidden because honesty about the work is the whole point.

Repo The question it asks The answer so far
Curvature-Grokking Does the commutator defect cause grokking? Yes — causally, and asymmetrically.
Exterior-Memory When does skew-symmetric wedge memory help? Not for recall (60/60 losses). Fine for aggregation.
Router-Drift Does explicit router supervision stop catastrophic forgetting in MoE? No. Diagnosed exactly why, proposed a partial fix.
TFM-Feature-Gap Do tabular foundation models close the feature-engineering gap? It depends on the dataset — verified against real Kaggle leaderboards.
Self-Calibrating Gradient Tuning Should every token get the same gradient signal? No. SGT reallocates per token, no proxy model needed.

DOIs

A flagship project exists. It's staying private until I reveal it.

What I stand for

  • Honesty about the work — floor effects, non-reproducing baselines, all of it goes in the paper.
  • Doing more with less compute — good science shouldn't require a data center. Good thing too.
  • Novelty — if something is already known, I want to know why it works.

Toolbox

  • ML: PyTorch, Kaggle (T4 ×2 — see: research philosophy)
  • Papers: LaTeX
  • Languages: Python · C · Luau (yes the Roblox one it's how the game dev started) · JavaScript · HTML · Bash
  • Currently learning: higher-level mathematics the kind my repos keep reaching for. The repos only cover the applied side so far.

Beyond ML

  • Game development in Luau
  • Front-end builds for my school
  • Cybersecurity
GitHub stats Streak stats

Pinned Loading

  1. Curvature-Grokking Curvature-Grokking Public

    Does commutator defect cause grokking? Causal experiments with auxiliary losses on modular arithmetic show amplifying it significantly impairs generalization; suppressing it helps.This is the main …

    Python

  2. Exterior-Memory Exterior-Memory Public

    Tests whether skew-symmetric (wedge-product) memory in sequence models is broken everywhere or only for associative recall. Replicates a known failure and shows it does not hold for permutation inv…

    TeX

  3. Self-Calibrating-Gradient-Tuning-for-LLMs Self-Calibrating-Gradient-Tuning-for-LLMs Public

    When you train a language model you are essentially teaching it to predict the next word. The standard approach is to treat every word equally during training. The model gets the same gradient sign…

    Python

  4. TFM-Feature-Gap TFM-Feature-Gap Public

    TeX

  5. Feonixid Feonixid Public

    This isnt a normal repo because in this repo we will be focusing on me Feonixid.Who i am my intrests what i stand for and where i want to be one day.

  6. Router-Drift Router-Drift Public

    Python