Research Notes
Documenting my path through LLM infrastructure, autonomous agents, production AI, and the questions that appear when the demos stop.
LLM Systems Deep Dive #7
Two papers published in 1991 disagreed on almost everything. The reactive versus deliberative debate set the terms for 35 years of multi-agent research.
LLM Systems Deep Dive #6
A budget of 500 tokens with the right memories can beat 4,000 with random ones. This is the architecture behind CogDB.
LLM Systems Deep Dive #5
Morocco is building an AI ecosystem from scratch. But sovereignty requires systems that work for the languages people actually speak.
LLM Systems Deep Dive #4
Software 2.0 changes the job from writing explicit instructions to writing specifications of intent. Our engineering tools have not caught up.
LLM Systems Deep Dive #3
Without temporal depth, a team of agents is a team of brilliant people with permanent amnesia. Here is how coordination frameworks address it.
LLM Systems Deep Dive #2
Two inference techniques attack the same bottleneck from different directions: one plays with time, the other with space.
LLM Systems Deep Dive #1
KV-cache fragmentation quietly limits how many users a GPU can serve. vLLM borrowed a 1960s operating-system idea to fix it.
Medium
The final chapter of a three-part journey that started with a dataset and ended with a fine-tuned language model anyone can use.
Medium
The first model memorized everything. The second attempt became a useful workflow generator. This is what changed between them.
Medium
A journey from curiosity to a production marketplace analyzer, with data on integrations, structure, and real automation demand.