Why
Memexponent was created to provide a nexus for system building explorations independently developed by Brian M. Dennis, PHD, a.k.a. bmdphd. For many decades now, I have also appeared across The Internets as crossjam.
Getting laid off in August 2025 (collateral damage of US federal research and development funding cuts) left me with a bit of free time for deep diving into some technical areas that had long been on my mind. Broadly construed these are in the area of AI Engineering, building applications using modern generative AI capabilities.
As well, I wanted to create a public project portfolio that I can freely share in prospective employment engagements. Something of my own unlocked from past jobs.
Themes
What even is AI Engineering? Chip Huyen has a pretty good definition and book (buy it!).
This book discusses AI engineering: the process of building applications with readily available foundation models. It introduces a practical framework for developing an AI application and efficiently deploying it.
It also covers how to navigate the AI landscape, including models, datasets, evaluation benchmarks, and the seemingly infinite number of use cases and application patterns.
AI Engineering a ridiculously broad space so here a few specific themes I’d like to emerge as I work on these projects.
Agentic Coding. AI seems promising as a tool for increasing developer productivity. There’s not a large body of rigorous work pointing pro or con. I’m skeptically optimistic based upon my reading of the technology media landscape. The best way to gather more evidence is to dive in and build things using tools like, aider, Claude Code, Zed, etc. So that’s what I’m doing. Man in the AIrena.
Retrieval Augmented Generation (RAG). Supplementing LLM based systems with information retrieval capabilities was a red hot engine of the hype cycle a year or two ago. The hype’s cooled off a bit, other than proclamations that “Rag is Dead”, but pragmatic industry knowledge seems to be maturing. Information retrieval is just the fancy academic way to say “search” and that’s long been a latent interest of mine, just from a pure systems perspective.
Systematic AI Evals. Before getting dismissed, I spent some time researching approaches to evaluating Generative AI systems. There’s an ever expanding galaxy of academic benchmarks and leaderboards for assessment. In addition, there are many different approaches to drive these systems, collect inputs and outputs, and calculate evaluation metrics. I’m a huge exponent of work from Eugene Yan, Hamel Husain, and others in this area. Always on the lookout for similar creators.
Context Engineering. This concept has only gained tractions over the last few months but it definitely resonates. Where I used to be dismissive of prompt engineering, I have now come around to taking seriously the technical challenges of getting an LLM good context. Phil Schmid has a good overview:
Context Engineering is the discipline of designing and building dynamic systems that provides the right information and tools, in the right format, at the right time, to give a LLM everything it needs to accomplish a task.
I have a hunch that some principles and techniques from the compiler community can provide a lot of leverage in this space.
Improved Software Engineering. Even though I’ve been at it for a while, I wouldn’t consider myself a “professional” developer. I’ve written lots of research code but haven’t set foot near a production system. This is an opportunity to grow my abilities in a number of software engineering practices, like testing for example.
Projects
retrocast is AI-assisted retrospective exploration of podcast episode archives. I have been subscribed to a decent number of podcasts for a solid decade now. Often times a new technology emerges as a theme across multiple podcasts. I then have questions like, “where else in my space of episodes did this concept appear?”, and “how long ago was the first appearance?”. I also have questions about persons, places, and things from a longitudinal perspective, such as “whatever happened to that company FooScape?”.
feedscope could be construed as the RSS feed version of retrocast. This project is more about pushing as fast and as far with agentic coding as possible. The basis is an extremely well documented API for the Feedbin aggregator service. How
prompthound is a pluralistic approach to examining Cory Doctorow’s blogging for the discerning craphound. Cory is a renowned author, activist, and blogger. After leaving BoingBoing and starting Pluralistic all of his blog content is readily available in full via RSS feed. A bit of a fanboi expedition, here I’m creating innovative search toys on top of this text dense, link rich, longitudinal, violently “of the Web” content.
Technologies
A smorgasbord of stuff that I either use routinely or have definitie plans for engaging with soon.
LLM Developer Toolkitss. llm, PydanticAI,
LLM Aggregator APIs. OpenRouter, TogetherAI, FireworksAI, et. al.
Local Model Platforms. LMStudio, ollama, llamafile, mlx
Model Context Protocol (MCP). modelcontextprotocol.io
Eval frameworks. Pydantic Evals, promptfoo, Inspect, MLflow Evaluation
MLOps Observability Frameworks. MLFlow, Pydantic Logfire
Marimo Notebooks. marimo feels like a well timed branching of Jupyter principles for the modern AI moment.
Coiled. Coiled is a nice platform for scaling Python computations on top of public clouds.
Hugging Face. Hugging Face is arguably the premier ecosystem for sharing LLM artifacts such as model checkpoints and datasets. They also provide model aggregation APIs, serverless inference, and low code ui deployment.
Terminal CLI AI Repls. Investigating what’s out there but I have this odd itch about an extensible REPL, combined with Agentic frameworks and MCP as a fruitful playground for high velocity experimentation with LLMs.
Etymology
A memex (a portmanteau of “memory” and “index”) is a hypothetical electromechanical device for interacting with microform documents and described in Vannevar Bush’s 1945 article “As We May Think”. Bush envisioned the memex as a device in which individuals would compress and store all of their books, records, and communications, “mechanized so that it may be consulted with exceeding speed and flexibility”. The individual was supposed to use the memex as an automatic personal filing system, making the memex “an enlarged intimate supplement to his memory”.
exponent (plural exponents)
- One who expounds, represents or advocates.
- (mathematics) The number by which a value (called the base) is said to be raised to a power in exponentiation. Synonym: power
memexponent
- A hypothetical value by which the power of a memex is exponentiated.
- An advocate of The Memex Method
- A portmanteauing of a portmanteau.