Building AI systems from scratch (RAG, agents, memory)
Web dev → AI engineering
Sharing what I learn while building
DMs open — always down to talk RAG/agents/AI
Finetuned Llama-3.2-3B-Instruct to reason like DeepSeek R1.
4bit quant. ServiceNow R1 dataset. Runs local and on Colab.
The model now thinks step by step, that wasn't there before I trained it.
Unsloth made it easy. No GPU pain.
Still learning. This one felt real :)
#ai
Most agents drown in stale tool outputs, ballooning token counts without adding value.
Compress & Prune drops those outputs and summarizes history before each step, cutting usage by 84%.
What's the one thing your agent remembers that it really shouldn't🤔
#Ai#llm#agent
Stop using the context window as a database;
build an external memory loop.
Write to a vector store, then select chunks via tool calls to prevent poisoning and speed the loop.
What happens when your agent pulls memory only when needed?
More tokens won’t fix your AI agent’s limits || they’ll clutter its reasoning.
Flooding a 200k-token window with tool outputs creates Context Distraction.
Trim the excess, keep prompts tight, and let the model focus.
What will you cut to keep your agent sharp?
#agent#context