20266 posts

1. November

   • 09Fixing the RNN's first draftWhy I'm not retraining the Brain-to-Text baseline, and treating its decodes as a rough draft for an online n-gram model to correct instead.
   • 09Mocking the metric firstFor the Helios corn climate challenge I rebuilt the scoring metric before any features, so I could validate offline instead of spending submissions.
   • 09A farmer that only plants wheatKaggriculture's first agent is a stub that plants wheat everywhere. Here's the modular design around it, and why heuristics and bandits come before RL.

2. August

   • 01Action space: MCP serversHow Tom reaches external platforms and agents outside his local hub, by putting their MCP servers on shared subnets.

3. January

   • 27Encoding observation statesScoping key-value retrieval for Gaby's memory, and choosing a lazy dtype-pattern hash over vector storage for his temporal lobe.
   • 01My playground setup and memorial wallThe platforms I use most for serving, training and sandboxing models, plus the lessons I keep relearning.

202523 posts

1. October

   • 24AWS AI Agent Global HackathonEntering Gaby as an autonomous data-cleaning agent on AWS Bedrock AgentCore and Sagemaker, and pausing competitions afterwards to rethink security.
   • 20BigQuery AI hackathon: what went wrongEntering Gaby into Kaggle's BigQuery AI competition, a week of refactoring, a banned Google Cloud account, and what I'd do differently.
   • 05Data quality, assurance and insightsThe eight data quality layers Gaby checks, assurance for regulated domains, and the default dashboard pages he builds.
   • 01Gaby's architecture and integrationsThe MCP action space, FastAPI backend, 2025 cloud services comparison, service checklist and integration partner tracker behind Gaby.

2. September

   • 20AI swamp: designing smarter Python agents before they mutateTwo OOP designs for agent base classes that adapt to different model hosting platforms.
   • 18Stages IV and V: anomalies and housekeepingOutlier detection during EDA, then the end of the pipeline: PCA-reduced subsets, end reports and what the user sees next.
   • 15Serving LLMs with Ollama (and why I archived it)A two-container Docker Compose setup for FastAPI + Ollama, the trade-offs of self-hosting, and the move to Vertex AI.
   • 14Stage II: missing dataDiagnosing MCAR, MAR and MNAR missingness per data field, a decision tree for it, and the agents that pick a handling strategy.
   • 10Stage I: defining and understanding the datasetWhat makes data unclean, why LLMs label data types inconsistently, and swapping prompts for sentence-transformers and cosine similarity.
   • 08Gaby's data-cleaning pipelineThe objective, user inputs and five general stages of Gaby's self-orchestrated data-cleaning workflow.
   • 06RL and temporal isolationWhy Gaby learns with numerical methods instead of fine-tuning LLMs, the pros and cons of isolating his memory, and why cognitive modelling got shelved.
   • 05Planning Gaby's decision makingFine-tune him like a reasoning model, or corner him with test cases in a little SIM world? Both methods, side by side.

3. August

   • 22How has this not been automated?The real-life problems behind Gaby: biased text labels, Monday-morning pipeline crashes, and botnets that live a little longer on good days.
   • 20Who is Gaby?Tom's database gatekeeper, a self-maintained data workflow with a PageRank-inspired model that decides which memories go where.
   • 03Tom's Social Circle: Retaining Data from External SourcesConnecting agents over MCP, and keeping Gaby as the gatekeeper for Tom's temporal and long-term memory.

4. July

   • 22Designing a game: a simple chatbotMulti-session chatrooms for seekers waiting on a reading, broken down into small hooks with Socket.IO and React.
   • 13#20“Everyone knew it was impossible, until a fool who didn’t know came along and did it.”

5. June

   • 10How I read the cardsSketching down the framework behind my own tarot predictions, one spread worked all the way through, before asking an LLM to do the same.

6. May

   • 04Attention as a ranking model, from scratchA small multi-head attention ranker meant to sit between a vector store and an LLM as a memory gatekeeper, trained on synthetic data with ordinal targets.
   • 02Tom's Memory NetworksDefining Tom's memory before building his decision-making components, and which model replicates short-term memory best.

7. February

   • 24My Role ModelsMy father, my mother, and the Yin and Yang they left me to balance.
   • 14Can a machine have emotions?Three ways to map a language model's weight space to emotion, and why I went with decoding parameters.

8. January

   • 05To our creatorA letter to the creator of our universe, from someone trying to build a human in code.

202420 posts

1. December

   • 06Cloning myself into a Language ModelQuick update on Tom's progress, adapters, and the policy function behind his responses.
   • 02Seconds beforeAs a child, I spent many days on my own and only ever peeked outside my window when my tummy made noise.
   • 02Treehouse: a pairwise dialogue classifierSorting incoming chat queries into one of two categories with sentence embeddings, and comparing Euclidean, Manhattan and cosine scores before choosing one.

2. November

   • 21Steering GPT-2 with a logits processorPlankton, a custom Hugging Face LogitsProcessor that boosts some tokens and anchors others at every decoding step, tested on "I love you. Do you love me?"
   • 19Looking inside GPT-2's attention blocksPlotting GPT-2's raw attention for "I love you" against "I hate you", block by block and head by head, to see where the two sentences start to differ.

3. October

   • 09To be a SuperhumanTo find hope amidst all chaos.

4. September

   • 19Lucky 30A letter to my future self, in case a machine is the one reading it.

5. August

   • 24Chasing my own tailLearning in this mortal life is like taking a bus ride home, stopping at every stop to pick up sticks and stones.

6. May

   • 29#9Every wonderer enjoys the state of oblivion.

7. March

   • 21To the end of all that I knowTo finally live the fantasy your younger self once dreamed of behind the counter at Taco Bell.

8. February

   • 07#8If every idea I have pondered exists on the internet, then surely someone else has implemented it.
   • 04Anne SmithAnne Smith visits the same shops every Saturday, and on Sunday, she attends church.
   • 04Reviving Hopfield Neural NetworksWhy a Hopfield network could be Tom's short-term memory, and where it stops being useful.

9. January

   • 29#7Let's play with Hilbert spaces as we wait for the quantum computing field to advance.
   • 14Lost in Nostalgia’s Spider WebsA high school flowchart, an afternoon detention, and the cobweb of probabilities that led me from algebra to AI.
   • 14#6If the worst outcome is simply giving up, then what do you have to lose?
   • 07#5If there is an afterlife, I hope to reincarnate as a ladybug.
   • 06#3Once you have mastered the foundational aspects of your endeavors, you can afford to be impatient.
   • 06#4Anything can be transformed into something greater when you have hope.
   • 02#2The worst time to doubt is when you are learning.

20234 posts

1. July

   • 26#1Wherever there is a need for improvement, there will also be a need for automation.
   • 10My favourite challengeTwenty minutes to solve what my colleagues solve in a few, and why I still love maths.

2. June

   • 28Am I biting more than I could chew?
   • 12Am I actually bad at maths?Short answer: Yes. Long answer: No. Well, at least I hope not.