Hi HN, I am shubham a 3d artist who learned coding in college as an I.T. graduate know logics but not an expert as i just wanna try my hands on to ai

So i built Resilient Workflow Sentinel this is offline ai agent which classify urgency (Low,Medium and HIgh) and dispatches to the candidates based on availability Well i want an offline system like a person can trust with its sensitive data to stay completely locally

Did use ai to code for speeding and cutting labor.

Its works on RTX 3080 system (this is an basic affordable setup not heavy ai machinery) which i want it to make it reliable without heavy upgrade This is full system doesn't require ollama(I am not against it)

I see in companies tickets are raised on jira and slack. Currently people or manager (self) have to sort those things either manually read one by one or send them to the cloud. But the issue is you can't send everything like there is a lot of sensitive data out there which they do not trust and makes it harder and manual sorting through thousands is likely a nightmare.

But then just imagine u get all the task classified like its urgency and distribution u can selectively see which task is urgent and needs immediate attention and last of all information doesn't leave your building totally secure Also Api sending is not the only issue u are paying per token cost for task for each may be monthly 100$ to 1000$ which can like save hassle for startup a lot or companies as well

There was several biases like positional bias also json out put bias also have issues in attention At start i tried just prompting things like Chain of thoughts,RISE(evaluate negative first), given negative examples,Positive examples, somewhere it was struggling with commonsense issue so examples for that (Later changed the approach)

Well prompting did give the output and worked well but took too much time to process for single task like 70 to 90secs for a task

Then i tried batching and the biases got worst like it got stronger it always use to like favour alice also more prompts are like ignored and more

For json output i used constrain so model can only generate json and if fails there is a as well parser i used when i implemented prompting only

This reduce time from 90sec to nearly 15 to 30secs per task I used steering vector to correct the attention i seen issues happening

Stack: Language: Python 3.10 Model: qwen2.5-7b-instruct Libraries: Pytorch, Hugging Face Transformers (No Langchain, No Ollama) API: Fast API UI: NiceGUI Hardware: Ryzen 5, 16Gb ram RTX 3080

Implementation:

Quantization: Load model in nf4 quantization so models like 7b can fit on vram of 10gb which is on rtx 3080 also my hardware

Steering Vectors: Standard prompting wasn't enough. I need to block or direct certain things on a certain layer of llm to make it reliable.

Json Constraints: Used constraint to make model strictly give json and also stop from over explanation this happens at logits level where token are blocked which are not required etc

github : https://github.com/resilientworkflowsentinel/resilient-workf...

Youtube: https://youtu.be/tky3eURLzWo

Amazon expects capex to hit $200B in 2026

Show HN: Local task classifier and dispatcher on RTX 3080(No Ollama, raw Python)

Amazon Falls After Vowing to Spend $200B on AI This Year

Show HN: Cryptographic toolkit in a single HTML file

Marketplace to buy/sell cheap Claude credits

The largest life-form on land 400M years ago scientists can't explain

The RCE that AMD won't fix

Paramount shuts down TV network available in major US cities without warning

Starlink Blocks Russian Troops' Internet Access, at Ukraine's Request

Please stop using OpenClaw, formerly known as Moltbot

The Melania coin is down 98% since its launch

India: Supreme Court Questions Meta and WhatsApp over Privacy Policy

Show HN: Askie – Child-safe AI voice chat and art creation for kids (ages 4-15)

Crypto Takes a Deep Slide Despite Trump's Support

Grafana Assistant gets playbooks: agent guidelines and commands in one

TSMC invests 17B for 3-nanometre fab in Japan

Lectures in London

Show HN: Calfkit – an SDK to build distributed, event-driven AI agents

Trump administration to end job protections for up to 50k federal workers

Ask HN: Those holding large amounts of software stock, how much have you lost?

AutoOverlap: Enabling Fine-Grained Overlap of Computation and Communication

Data Science Weekly – Issue 637

U.S. Dealers in Full Panic Mode After Canada Green-Lights Chinese Cars

Sidestepping Evaluation Awareness and Anticipating Misalignment with Evaluations

Show HN: Ecommerce CFO Skill – Diagnostic Frameworks for AI Agents

Cubans rendered powerless as outages persist and tensions with US escalate

The unfathomable Minnesota transcript that tells the reality of America today

Show HN: DoAnything – AI agent with its own email/phone for your to-do list

I've spent 20 years as a touring guitarist. The live industry is broken

Why a Fintech Unicorn Is Donating Millions to Start a Credit Union