Aswin Kumar ss

CS undergrad building low-level systems & AI infra in Rust & C. Augmenting POSIX shell with async IPC, LanceDB vector memory & llama.cpp.

[Aswin Kumar ss @SsAswin12481]

Part 4: This is bug fixing of part 3 ans also show the dynamic model changing. So what happened last time was my 7b model answer according to the context itself but my 1.5b model didn't So I researched the root cause and found a way ⬇️

[Aswin Kumar ss @SsAswin12481]

@Raigeki_Dev It’s built from zero using pure C and POSIX system calls, talking directly to the Linux kernel. I even wrote a custom Rust orchestrator to manage GPU VRAM for local AI. Check out the pinned repo on my profile if you want to see the source code and updates 😊

[Aswin Kumar ss @SsAswin12481]

@Raigeki_Dev My shell is not a wrapper I made the the whole shell my own

[Aswin Kumar ss @SsAswin12481]

@Raigeki_Dev @X Lets connect😊

[Aswin Kumar ss @SsAswin12481]

@Officially_Dev_ For me legion and I am also using that because I want that local models to run

[Aswin Kumar ss @SsAswin12481]

So found way is just make my rust engine decide check for memory in vector db if found then write it in chatML format and give it like this is the fix now what the model just tells that this will work only if it have high confidence or less then model gives the answer

[Aswin Kumar ss @SsAswin12481]

Part 4: This is bug fixing of part 3 ans also show the dynamic model changing. So what happened last time was my 7b model answer according to the context itself but my 1.5b model didn't So I researched the root cause and found a way ⬇️

[Aswin Kumar ss @SsAswin12481]

So SLMs either they are trained to give to do tasks or fixes and all they dont have that much of intelligence compared to 7B models and higher so when I researched I found about OCC-Rag models so they are best to answer on basis of context and only 1.7b.

[Aswin Kumar ss @SsAswin12481]

But then the problem is it only answer on context we need answers too so then i found was cad context aware decoding but then that's just introduce more computing and the problem if the model knows better answer that's just ignored.

[Aswin Kumar ss @SsAswin12481]

Oki the model fix i will check what happened 1.5b next task is that I feel rag is finding due to shell errors are same when lsh not found so rag fetched from lancedb the prev error.

[Aswin Kumar ss @SsAswin12481]

Part 3 of building my low-level C/Rust AI shell (aish): Built an automated Memory Consolidator.

Instead of just storing stateless chat history, I built a two-stage memory pipeline. also did dynamic modelling 1.5b for error 7b for chat 14b for thinking detail explanation below

[Aswin Kumar ss @SsAswin12481]

for validation and checks for all episodic memory is there was a fix yes promote that all. For dynamic model sorry i didnt showcase the 14b model one I will post taht soon its working and you can also see some bug with error suggestion like that only came while recording.Iwillfix

[Aswin Kumar ss @SsAswin12481]

The diff is Python used ollama so local server and where as rust used llama.cpp serverless the model caching and using gpu in full extend was attained by using the rust. Both used same model Qwen 7B

[Aswin Kumar ss @SsAswin12481]

Part 2 of building my low-level POSIX shell (aish): Replaced the slow Python backend with an async Rust daemon.

The Python prototype took 18.6 seconds to respond. Rewrote the AI orchestrator layer in Rust using IPC over Unix domain sockets.
The result 2.0 secs completion

[Aswin Kumar ss @SsAswin12481]

I had some indexing on python that's why its answering on basis of my shell. After this I will integrate a memory engine for this rust backend on part 3 :)

[Aswin Kumar ss @SsAswin12481]

@senodev_ Hi lets connect

[Aswin Kumar ss @SsAswin12481]

@md_kasif_uddin Github

[Aswin Kumar ss @SsAswin12481]

@senodev_ @X Lets connect