Skip to content

Working notes on AI systems and the software under them.

Language models, agents and the protocols wiring them to real tools. Linux, Python and the tooling underneath. Written up as I learn it, one post at a time.

Read the latestRead the latestBrowse all 47 postsNew here? Start with a path
EngineeringFeatured

Feb 24, 2025

Part 2 of 3 · The Model Context Protocol

Docker's Dynamic MCP and the Context Window Problem

Fifty tool definitions crammed into the context window costs tokens, degrades reasoning, and risks hallucinations. Docker's containerized MCP Gateway and dynamic discovery flip the model.

45 min read

Kushan Manahara

Machine Learning Engineer · personal engineering publication

articles
47
topics
7
reposts
0

Most viewed articles

What people actually read this year.

View allView all

Latest writing

Filter by what you came for.

View allView all
AllSystemsEngineeringCareerResearchInferenceEvaluationExperiments

Series

Longer arguments, split into parts.

All series
4parts

Moving to Linux

Picking the OS, understanding why the tooling favours it, then fixing the things that break first.

Read the series
  1. Why Programmers Keep Choosing Linux
  2. Choosing an Operating System for Programming
  3. How to Install MySQL on Kali Linux
  4. Fixing the Legacy trusted.gpg Keyring Warning on apt update
2parts

Regression, Explained

The two models everything else is built on, worked through from the maths outward.

Read the series
  1. Linear Regression Explained
  2. Logistic Regression Explained
3parts

The Model Context Protocol

What MCP solves, what happens to the context window at scale, and where it sits next to A2A.

Read the series
  1. Why I'm Excited About MCP
  2. Docker's Dynamic MCP and the Context Window Problem
  3. MCP and A2A Are Layers, Not Rivals
5parts

PyTorch

Building the framework up from its one real primitive — the tensor — through autograd and the training loop.

Read the series
  1. PyTorch: From Tensor to Neural Network
  2. Autograd: How PyTorch Learns
  3. nn.Module: Building Networks That Hold Together
  4. Datasets and DataLoaders: Keeping the GPU Fed
  5. Training for Real: GPUs, Mixed Precision, and Schedules

Topics

Things I keep coming back to.

All topicsAll topics
Systems
#1

Systems

10 posts

Engineering
#2

Engineering

18 posts

Career
#3

Career

5 posts

Research

Research

6 posts

Inference

Inference

3 posts

Evaluation

Evaluation

3 posts

Experiments

Experiments

2 posts

Join the mailing list

One email when a post goes out. No digests, no drip sequence, no sponsors.

  • 01New posts, and corrections to old ones
  • 02The occasional note on something I got wrong

Join the technical mailing list. Free updates, unsubscribe in one click.

Most read this year

01Getting Started with DjangoEngineering · 746 views02How to Stop Conda Activating the Base Environment AutomaticallyEngineering · 728 views03Fixing pip's externally-managed-environment ErrorEngineering · 684 views04How to Install MySQL on Kali LinuxSystems · 356 views05Why Programmers Keep Choosing LinuxSystems · 29 views

The archive

47 posts, four years, no reposts.

Everything is dated and versioned. When a finding turns out to be wrong, the correction sits on top of the original, not in place of it.

47

Articles published

7

Topics covered

Open the archiveOpen the archive