Anuj

cse'28 ug | full-stack • devops • gen-ai | open-source contributor
#buildinpublic

[Anuj @0xzdev]

@VrajTalati1 token budget. we count backward from the newest message until it hits 35% of the context window that's the rough cut point.if that cut point would split a tool call from its result, it moves forward to the next clean break(the next message that isn't part of a tool call pair)

[Anuj @0xzdev]

i built a minimal terminal coding harness

features :
-runs bash commands
-edits code inside a sandboxed workspace
-spawn subagent to explore the codebase
-does compaction to maintain context-window

demo video and repository link below

[Anuj @0xzdev]

@VrajTalati1 subagent output reaches the LLM like any other tool result. At 85% context, a 2nd LLM call summarizes the transcript into a handoff note, replacing everything up to a cut point but a recent tail stays as-is below it, unsummarized, so recent context stays exact.

[Anuj @0xzdev]

@X_Ibyte thnx !!

[Anuj @0xzdev]

@Avichal_08 thnx!!

[Anuj @0xzdev]

repository link :

https://t.co/xFaePqKKcS

[Anuj @0xzdev]

@WarBornePhoenix thnx !!

[Anuj @0xzdev]

built a load balancer in GO

features :
1. concurrent request handling
3. reverse proxy
2. fault detection
3. health checks
4. thread safe shared state
5. pluggable algorithm selection using ocp

demo video and repository link below:

[Anuj @0xzdev]

@_mishkyn sure, i will load test it

[Anuj @0xzdev]

@Amitesh48256 :)

[Anuj @0xzdev]

@Avichal_08 thnx !!

[Anuj @0xzdev]

repository link :

https://t.co/tQu8nKlLvC

[Anuj @0xzdev]

@GonnabeNikhil i built an mvp to simplify the process of understanding and analyzing large code bases
https://t.co/K8tNWPAKoP

check out the post and repo (still in progress..)

https://t.co/4tE9Dyi4ba

[Anuj @0xzdev]

I built Repolyzer ,an AI powered github repository analyzer which makes navigating large codebases easier
it helps you understand a codebase by generating:
-system architecture
-intelligent summaries
-context aware conversations for debugging and resolving issues

Demo video

[Anuj @0xzdev]

Ever wondered what actually happens after you hit enter on an LLM?

I wrote a breakdown of the journey of your input inside an LLM

https://t.co/qYra1RHdFZ

[Anuj @0xzdev]

repository link :

https://t.co/9JPRoMi0uv

star it if you enjoy !!

[Anuj @0xzdev]

I trained a neural network to play snake game live in your browser.
A Deep Q-Network written from scratch in TS. Press Start and watch it go from flailing to deliberately chasing food. The best part: it keeps training while you switch tabs.

live link : https://t.co/pHLaLfFFW3

[Anuj @0xzdev]

@Deep_Code_007 its just a fun project ,but to answer your question actually the bias accumulates over the generations due to non zero mean ,which leads to sigmoid saturation ,and the network stops responding ,also the pipe velocity is increasing every second so after 40 sec it becomes too fast

[Anuj @0xzdev]

Made Flappy Bird where nobody plays it the birds train themselves
A population of neural networks learns to fly via a genetic algorithm

wrote it from scratch in TS

Here is the demo video and live link :
https://t.co/c6b6ewruVW

repo link is below

[Anuj @0xzdev]

@garvitsinglaa basics of genetic algo and NN

[Anuj @0xzdev]

@aashaykulk its randomized , the gap height is rerolled

[Anuj @0xzdev]

@AmanBuilds_ we put ~3000 birds in the game ,each bird decides when to flap ,birds that survive the longest become parent and makes the next generation ,and the newer generation are better at the game