Seed
Agentic System Design Checklist
Before drawing boxes, pin down the decision boundary, actions, tools, guardrails and memory; then design the architecture and plan for latency, hallucination, scale and integration.
AI, machine learning
$ cd ./topics/ai-ml
AI, machine learning & research
Machine learning fundamentals, AI agents, and the research I did at university: topic modelling, social network analysis and online communities.
26 notes🌱 14 seed🌿 12 growing🌳 0 evergreen
I specialised in AI and machine learning at university, and my dissertation used topic modelling and social network analysis to study online communities. These notes collect the fundamentals, that research, and how I think about AI agents now that they write a lot of the code.
Start here
- Machine Learning · Deep Learning · Reinforcement Learning · Natural Language Processing
- What Makes an AI Agent · Don’t Read the Code? Verifying AI-Written Code
Research
- Pathways into Cybercrime, my dissertation
- Topic Modelling with LDA · Social Network Analysis · Mining Reddit for Research
- Online Communities and Interaction Spaces · Hacking Forums as Learning Communities · Why People Legitimise Hacking
Open questions
- What should a platform give an AI agent by default, and what should it never let one do?
- How do you verify work you no longer read line by line?
- Agentic Development Platforms
- Agentic System Design Checklist
- AI Before the LLM Boom: Notes from 2019
- Blue-Green Deployment
- Books That Shaped How I Build
- Building a Graph Component Library
- Building a Web Platform for 80 Sites
- ChainBuddy - Two-Phase Agents and Model Tiering
- Cloud Platform Translators
- Contract-Driven Development
- Convex
- Deep Learning
- Designing a Platform for AI Assistants
- Don't Read the Code? Verifying AI-Written Code
hover to explore · click to open
Notes (26)
- Seed
ChainBuddy - Two-Phase Agents and Model Tiering
ChainBuddy solves the blank-page problem with two phases: a cheap, fast conversational agent turns a vague request into a spec, then a planner/worker/connector backend builds the LLM-evaluation pipeline.
AI, machine learning - Growing
Eureka: Automating Reward Design with Coding LLMs
NVIDIA's Eureka puts an LLM in a loop around reinforcement learning — it reads the environment code, writes reward functions, trains, reflects on the results and rewrites. A blueprint for closed-loop agents.
AI, machine learning - Seed
Evaluating AI-Assisted Work - Cognitive Load and Over-Reliance
People are poor judges of how much an AI helped them, so measure AI tools from several angles (NASA TLX, usability, output quality, observed behaviour) and design so the human stays the final judge of quality.
AI, machine learning - Seed
Guardrails and Human Oversight for AI Agents
Layer an agent's defences, start with PII, safety and irreversible actions, and choose deliberately where the human sits: in the loop, on the loop or above it.
AI, machine learning - Seed
MACRS - Multi-Agent Conversational Recommender
MACRS is a conversational recommender built from several agents that reflects on failed recommendations at two levels, simulates users with another agent, and beats single-agent systems. A good lesson in optimising the system, not just the model.
AI, machine learning - Seed
Multimodal Web Agents
An agent that only reads HTML never sees the final render. Multimodal web agents like WebVoyager read screenshots like a person does, which unlocks real browsing and introduces the grounding problem.
AI, machine learning - Seed
ReAct vs Plan-and-Execute
ReAct decides one step at a time from what it just observed; plan-and-execute writes the whole plan first and then runs it. Choose per agent, not per system.
AI, machine learning - Seed
Smart Parking Multi-Agent System - Case Study
A worked design exercise: turning reactive parking microservices (sensors, booking, users, notifications) into a proactive manager-worker multi-agent system with driver, zone, sensor and reservation agents.
AI, machine learning - Seed
Agentic Development Platforms
AI agents don't just write code; they need infrastructure to run it. Solid internal developer platforms are the building blocks of agentic ones, and without them organisations fall into one of three traps: blocked, ungoverned or bottlenecked.
Platform engineering - Growing
AI Before the LLM Boom: Notes from 2019
In 2019 I went to two industry events, an AI conference and Big Data LDN, and wrote up what I heard: data is the new oil, edge inference, AI for business users. Read again in 2026, some of it was prophetic and some of it was the wrong bet.
AI, machine learning - Growing
Designing a Platform for AI Assistants
A worked example: several teams each build AI assistants from scratch. What should a platform own, what stays with teams, what do you buy, and how do you win buy-in, discover features, plan, build and iterate?
Platform engineering - Growing
Image Analysis Fundamentals
My 2021 university notes on image processing, histograms and pattern recognition, distilled: images as intensity functions, contrast stretching, features and classes. The vocabulary underneath modern computer vision.
AI, machine learning - Seed
Don't Read the Code? Verifying AI-Written Code
The author of Clean Code now argues humans shouldn't review AI-written code line by line: specify what correct means, verify it with deterministic tools, and keep humans accountable for specs and sign-off.
Testing, quality - Seed
Deep Learning
Neural networks with many layers, trained end to end by backpropagation — CNNs for images, transformers for sequences, and the bridge from university neural nets to today's LLMs.
AI, machine learning - Seed
Hacking Forums as Learning Communities
Game hacking and modding forums are, first of all, places where people learn — sharing tools, tutorials and status. The same dynamics that make them great teachers raise the question of where the learning leads.
Learning systems - Growing
Machine Learning
Learning patterns from data instead of hand-coding rules — supervised, unsupervised and reinforcement learning, and the habit of testing on data the model has never seen.
AI, machine learning - Growing
Mining Reddit for Research
Reddit's topic-focused communities make it a rich research source. How to collect posts with PRAW and an archive, how to clean them, and the ethics that keep individuals out of your results.
AI, machine learning - Seed
Natural Language Processing
Turning human language into something machines can work with — tokens, vectors, embeddings — so they can classify, extract, summarise or generate it. From bag-of-words and sentiment lexicons to transformers.
AI, machine learning - Growing
Online Communities and Interaction Spaces
What actually makes an online community — shared purpose, interaction, norms, a platform — and why "member" and "participant" aren't the same thing. The literature behind how I framed forum research.
AI, machine learning - Growingproject
Pathways into Cybercrime
My final-year dissertation at Kent. Can topic modelling and network analysis of online communities show whether game hacking and cheating forums lead people towards more serious hacking? The question, the approach, the ethics, and what I'd do differently with today's LLMs.
AI, machine learning - Growing
Reinforcement Learning
An agent takes actions in an environment to maximise reward. The vocabulary — policy, reward, rollout — and why reward design is the hard part.
AI, machine learning - Growing
Social Network Analysis
Modelling people and their interactions as a graph to find who's influential, which communities form and how behaviour spreads. Centrality, clustering, Louvain community detection, and why graph thinking travels so well.
AI, machine learning - Growing
Topic Modelling with LDA
Latent Dirichlet Allocation treats each document as a mix of topics and each topic as a mix of words. It's unsupervised, fast to try, and still a useful explainable baseline next to embeddings and LLMs.
AI, machine learning - Growing
What Makes an AI Agent
A chatbot answers; an agent perceives a change, reasons about a goal, plans, and takes several actions with tools to get there.
Platform engineering - Seed
Why People Legitimise Hacking
People are more willing to see hacking as legitimate when they feel a system or authority won't listen to them, and they're angry about it. A psychology finding with uncomfortable lessons for organisations.
AI, machine learning
Coming up
- Agent-driven QA and platform dry runs experiment
- A camera-only capture app for paper notes experiment
- Personal AI agent apps explore
- Sketch-to-form builder experiment