JM

Field Notes

Field Notes

Field Notes is an ongoing technical journal documenting work at the intersection of creative technology applications, scientific communication, media systems, AI workflows, and digital infrastructure.

Some entries begin as small observations gathered in the middle of active projects. Others expand into long-form essays, architectural breakdowns, production notes, postmortems, or research experiments. The common thread is an interest in how systems are actually designed, operated, repaired, and adapted in practice.

Much of the material here comes from the space between strategy and execution: the translation layer where ideas encounter technical constraints, organizational realities, human behavior, and the occasional unexpected breakthrough. The goal is not simply to present finished outcomes, but to preserve the process behind them.

Showing 15 of 15 notes.

August 5, 2026

Dispatches from the Argument Factory

data-and-classificationautomationlocal-firstmedia-and-productionProject: rsrag 5 min read Dispatches from the Argument Factory This project began as a joke about making Sam Altman and Dario Amodei fight as badly coordinated robot boxers, but it quickly turned into something much more serious: an experiment in open models, local AI, comparative retrieval, and evidence-grounded argumentation. Frustration with expensive, opaque cloud tools pushed the work back toward inspectable local systems, while the original boxing gag evolved into a platform that can build curated intellectual corpora, uncover hidden disagreements inside a user’s question, and measure how much the written record actually supports the fight. The project was built across several AI collaborators, each seeing a different part of the work, so the article ends by letting those models tell the story from their own corners.
July 18, 2026

OpenAI Build Week: What Would You Ask Sam?

applied-aiweb-developmentProject: samnavsemantic-searchai-workflowsstart-here 4 min read OpenAI Build Week: What Would You Ask Sam? OpenAI emailed me about Build Week with four days left, which meant choosing between several questionable Codex projects and figuring out how to make a meaningful addition to SamNav—an application that was already built, deployed, and working. The answer became “What would you ask Sam?”: a question-first way to navigate Sam Altman’s public writing without pretending to impersonate him. The new system uses SamNav’s existing taxonomy, semantic dossiers, and GPT-5.6 to interpret nuanced questions and plot a grounded route through things he has actually published.
June 21, 2026

The House Always Eats

culture-and-critiqueweb-developmentProject: sandwich-betssocial-softwarenonsensesystems-designanti-extraction 9 min read The House Always Eats Sandwich Bets starts with a stupid premise and follows it to its logical conclusion: what if the only safe wager left is lunch? Sandwich Bets becomes a napkin ledger turned into software: a way to remember the terms, preserve the obligation, and cap the damage at lunch. From there, the essay widens into a critique of gambling culture and the casino internet, where platforms manufacture suckers by turning human instincts for stakes, prediction, rivalry, and being right in public into habit and revenue. Sandwich Bets keeps the ritual but refuses the extraction: no cash, no casino, no wallet, one-sandwich max. The deeper turn is that it is less a betting app than a gifting system in disguise. Being wrong never felt so right.
June 4, 2026

Navigation for the End Times

web-developmentProject: samnavweb-designclassificationanti-extractionnonsensestart-here 6 min read Navigation for the End Times SamNav is a tiny independent index for Sam Altman’s blog, built because the posts are worth reading but the navigation is almost comically minimal. It adds search, topics, reading paths, related posts, and an admittedly dry Navigation Severity Score without republishing the original writing. It’s part joke, part web repair: a small Dockerized map for a corner of the internet that needed one.
May 30, 2026

Expertise Without Permission

culture-and-critiqueapplied-ailabor-and-classcraftautomationanti-extractionnonsensestart-here 9 min read Expertise Without Permission A working-class reflection on AI, expertise, and gatekeeping. The piece argues that expertise is real, but access to it has always been shaped by class, credentials, and permission. AI disrupts that arrangement by letting more people turn ideas into practice before institutions decide they are “qualified.” That democratizing potential is real, but complicated: the same tools that weaken old gates are mostly owned by corporations, turning liberation into subscription-based dependency. The essay defends craft while rejecting the fantasy that craft alone protects workers.
May 22, 2026

A Machine for Cross-Examining My Greed

applied-aiweb-developmentProject: signal-deskresponsible-webai-workflowsdata-architecturenonsensestart-here 6 min read A Machine for Cross-Examining My Greed Signal Desk is a local AI-powered investment research desk I built precisely because I don’t know enough about investing to trust my own instincts. It turns scattered curiosity about AI companies, infrastructure, robotics, sensors, and market hype into a slower research ritual: filings, thesis cards, source-linked briefs, recommendation categories, discovery scans, and ghost trades that test ideas without risking real money. It does not give financial advice or place trades. Instead, it acts as a contradiction engine, helping me separate evidence from narrative before I do something financially...
May 19, 2026

Before the Dream Becomes a Story

language-and-minddreamsmemorysymbolic-language 10 min read Before the Dream Becomes a Story This essay argues that dreams are often lost because we ask them to become stories too soon. In the fragile space after waking, dream memory rarely arrives as narrative; it appears as fragments: a color, a room, a body sensation, a person, a mood, a wrongness. Rather than treating those fragments as incomplete failures, the piece suggests they may be the first language dreams offer us: small, charged keys that can reopen a larger field of memory later. By treating words as handles rather than explanations, dream recall becomes less about recording a finished story and more about preserving enough shape for the dream to return.
May 17, 2026

The Ghost Switcher

media-and-productionapplied-aiProject: autonomous-producerlive-productionsystems-designlocal-firstagentic-systemshuman-in-the-loopstart-here 10 min read The Ghost Switcher Part three closes the series by moving from prototype to stakes: a real venue-network failure in London where the remote producer was disconnected, the stream dropped, and the online audience was left staring into the void. From that failure, the article argues that Autonomous Producer is not just about smarter switching, but about removing fragile remote-control dependencies from live production. The answer is a local “ghost switcher” that observes the run of show, source states, transcript cues, and device health; recommends actions in shadow mode; logs every decision; moves through human-confirmed switching; and only earns narrow autopilot after repeated evaluation. Its intelligence is not one giant model, but a resilient local control loop built from constrained commands, structured observations, production rules, and the discipline to hold when the evidence is bad.
May 17, 2026

The Show Has a Spine

media-and-productionapplied-aiProject: autonomous-producerlive-productionconference-techlocal-firstagentic-systemshuman-in-the-loopstart-here 9 min read The Show Has a Spine Part two follows the project from portable livestream workflow into live system architecture, opening with a makeshift hotel-room lab in Killarney where OBS, ATEM control, testing sources, and early automation experiments began turning into something more coherent. The central idea is that an autonomous producer should not be one giant AI model “watching” a show, but a local orchestration system that understands the show through a run of show, device state, sampled video, rolling transcripts, source-state signals, and a director policy. This part argues that scientific livestreams are not just sequences of cuts, but sequences of expectations, cues, timing, identities, and drift, and that the real technical challenge is giving the machine a structured point of view about what is happening before asking it to recommend or execute production decisions.
May 16, 2026

Notes on the Yellow Pelican

media-and-productionapplied-aiProject: autonomous-producerlive-productionconference-techlocal-firstagentic-systemshuman-in-the-loop 7 min read Notes on the Yellow Pelican Part one of this three-part series traces Autonomous Producer back to its practical origin: portable livestream kits built as conferences returned in person while online access remained essential. What began as yellow cases, hand-cut foam, simple signal paths, and laminated diagrams evolved into a more complex workflow involving cameras, OBS, ATEM controls, remote producers, and the latency of operating live shows over a network. The article argues that scientific livestreams often follow a repeatable grammar, and that Autonomous Producer is not a general AI director, but a local system that can observe the show, follow the run of show, execute constrained commands, log decisions, and gradually earn narrow autonomy.
May 15, 2026

Rebuilding My Portfolio With AI, Static HTML, and Stubborn Taste

web-developmentapplied-aiProject: portfolio-rebuildweb-designai-workflowscraftlocal-first 10 min read Rebuilding My Portfolio With AI, Static HTML, and Stubborn Taste The original version of my portfolio was built in 2021 from downloaded themes, hand-edited HTML, layered CSS overrides, and an unreasonable number of animated GIFs. The visual identity still felt right, but the underlying structure had become difficult to maintain and extend. This rebuild became an experiment in preservation through reconstruction: using AI-assisted workflows, static publishing, structured content, local admin tooling, and a lot of small human judgments to modernize the system underneath the site without sanding off the parts that still felt personal.
March 21, 2026

Signal to Strategy: From Scientific Activity to Demand Intelligence

data-and-classificationProject: signal-to-strategybusiness-intelligenceconference-planningclassificationlife-sciencesstart-here 4 min read Signal to Strategy: From Scientific Activity to Demand Intelligence Scientific activity is not the same thing as scientific demand. This article argues that publications, grants, registrations, abstracts, and engagement metrics only become strategically useful when interpreted together, classified consistently, and understood in context. Scientific demand intelligence is presented as a decision-support framework for identifying where research momentum is building before it becomes obvious through late-stage operational signals. Rather than replacing expert judgment, it gives organizations a more structured way to recognize weak signals, understand emerging communities, and make better decisions about what to convene, where to invest, and when to act.
March 20, 2026

Signal to Strategy: Why Scientific Classification is More than a Labeling Exercise

data-and-classificationProject: signal-to-strategybusiness-intelligenceconference-planningdata-architecturesemantic-searchstart-here 4 min read Signal to Strategy: Why Scientific Classification is More than a Labeling Exercise Scientific classification is more than tagging content after the fact. This article argues that classification should function as an interpretive layer inside the data architecture, giving messy scientific and operational records a shared vocabulary that makes them comparable, traceable, and strategically useful. Publications, grants, abstracts, investigator activity, and event signals all describe scientific activity from different angles, but without a durable classification layer they remain difficult to connect. By using Snowflake as the analytical backbone, supported by local validation workflows involving Postgres, Qdrant, OpenAlex, embeddings, and structured JSON outputs, the goal is to create an enrichment system that is inspectable enough to trust and flexible enough to refine. In a conference context, that layer can help identify emerging areas, converging communities, and better evidence for planning decisions.
March 19, 2026

Signal to Strategy: Why Scientific Demand is Difficult to Detect

data-and-classificationProject: signal-to-strategybusiness-intelligenceconference-planningclassificationlife-sciencesstart-here 3 min read Signal to Strategy: Why Scientific Demand is Difficult to Detect Scientific demand is difficult to detect because it rarely appears as a single clean metric. By the time it shows up in registrations, abstracts, sponsor interest, or meeting-planning conversations, the most useful window for strategic interpretation may already have passed. This article argues that demand often emerges earlier through weak, distributed signals across publications, grants, collaborations, methods, disease areas, and shifting scientific attention. The challenge is not collecting more data, but building systems that can classify, connect, and interpret those signals in context so organizations can recognize emerging relevance sooner and make better decisions about which communities to convene and where to invest.