Ajit Ashwath

someone hire me

[Ajit Ashwath @introcapybara]

@films_and_life may i join

[Ajit Ashwath @introcapybara]

@komalpreet2809 which one?

[Ajit Ashwath @introcapybara]

starting with diffusion models. very math-heavy, but gotta do this.

[Ajit Ashwath @introcapybara]

@kartikktwt yes bro

[Ajit Ashwath @introcapybara]

had my first on-site interview in gurugram for a (so-called) tech role.

the hiring manager looked at my resume & written assessment, then said, "you're too much for this role. decide wisely. this isn't a complete tech company."

i took the advice. rejected myself and walked away

[Ajit Ashwath @introcapybara]

@kartikktwt lol how

[Ajit Ashwath @introcapybara]

@kartikktwt go for it man

[Ajit Ashwath @introcapybara]

@DailyNoud nolan while writing that "crucial scene"

[Ajit Ashwath @introcapybara]

what does jee mains rank gotta do here?

[Ajit Ashwath @introcapybara]

@iamshreytyagi 😂

[Ajit Ashwath @introcapybara]

@kartikktwt 😂

[Ajit Ashwath @introcapybara]

so far, this is the most wtf question i’ve come across on lc

[Ajit Ashwath @introcapybara]

@keepingIt100a you gotta run the backend seperately

[Ajit Ashwath @introcapybara]

built a prompt compression engine that reduces llm prompts by up to 70% while preserving semantic meaning.

it uses textrank + bart for compression, minilm for similarity scoring, and tiktoken to measure token and api cost savings.

[Ajit Ashwath @introcapybara]

github: https://t.co/hRl3Jinm7d

[Ajit Ashwath @introcapybara]

@kartikktwt honestly a bad thing to do

[Ajit Ashwath @introcapybara]

@kartikktwt yoo 1.5 b viewed

[Ajit Ashwath @introcapybara]

@kartikktwt go to mohali nig

[Ajit Ashwath @introcapybara]

one interesting result from the regularization
increasing γ aggressively reduced tree complexity.

γ = 0 → ~6 leaves/tree
γ = 1 → ~3 leaves/tree
γ = 5 → ~1.5 leaves/tree

yet auc remained around 0.992–0.994.

[Ajit Ashwath @introcapybara]

implementing xgboost from scratch

tested on the breast cancer wisconsin dataset:
• 569 samples
• 30 features
• 62.7% benign
• 455 train / 114 test

scratch implementation:
auc: 0.9937
accuracy: 94.74%
logloss: 0.0983
avg leaves/tree: 4.3

[Ajit Ashwath @introcapybara]

i also ran 5-fold stratified cross-validation:

fold aucs:
0.9990
0.9889
0.9897
0.9980
0.9936

mean auc: 0.9939 ± 0.0041