fkx · Newton, Massachusetts
Making AI do new things. Conversational coding, and AI-native software development.
I build software by talking to it. Not autocomplete, not a copilot at the edge of an IDE - the whole lifecycle, conducted in conversation across several AI surfaces at once. That practice has a name and a shape, and most of what is below exists to test whether it actually holds up.
DIDMO — Designed, Implemented, Deployed, Maintained, Operated - by AI
A philosophy needs a mechanism or it is just a slogan. DIDMO's mechanism is a mesh.
Normally every stage of a software lifecycle is a separate infrastructure: build is CI, distribute is an app store or package registry, update is an auto-updater or MDM, operate is a cloud runtime plus monitoring. Four or five stitched-together systems, each with its own identity model, its own failure modes, its own seams. On a fabric where every participant already has identity and is already reachable, those stages collapse into one channel. The gain is not that each stage runs faster. It is that the seams disappear.
Designed
Distributed design across agent contexts.
Implemented
Distributed build - breaks the single-context ceiling.
Deployed
Enrollment is distribution. The membership channel is the delivery channel.
Maintained
Updates and health telemetry ride the same channel.
Operatednewest
Runtime fabric for genuinely distributed applications.
Operated was the weak letter for most of DIDMO's life - well-grounded in theory, thin in practice, and entirely dependent on the application actually being distributed. It only closed when a capable model began running locally on fleet hardware. Saying which letter was thin is what should make the other four believable.
A private compute fabric with a control plane I own - self-hosted coordinator, self-hosted relay, no dependency on any mesh vendor's service. Four GPUs across two operating systems addressable as one pool.
Federated training across the full fleet converges monotonically over three rounds, with raw data never leaving the node that holds it. The coordinator survived a full reboot of the machine it used to live on.
A verification layer for federal compliance regimes. Agents draft an answer from primary sources, adversary agents attack it, every surviving claim is tied to a citation and dated. Humans read; humans do not post.
Three communities over one shared authority graph, so a single regulation cited by two regimes reconciles into one view instead of two contradictory ones. I audited my own verification product in public and published what it got wrong.
196 pages on making AI genuinely useful in ordinary work, for readers with no technical background. First in the AI-Powered Author Toolkit series. Kindle and paperback.
Published October 2025.
The book above, turned into a conversational companion - ask it questions, get answers grounded in the text, with the illustrations pulled through.
Built entirely by conversation. It is the method demonstrating itself on my own material.
Memory as a product surface rather than a feature. Conversation graphs, model chaining, variant groups, and a coordinator layer for reasoning across several models at once.
395 commits.
A curated secondary market for NVIDIA compute - datacenter, workstation and prosumer cards for AI training and inference. Buyer tables, a market scout, a research hub, and category guides from DGX Spark and Jetson Thor down to used 3090s. Runs on the eBay Partner Network.
Its organising principle is "start with the job, not the part number", which is the commercial expression of a measurement: I benchmarked six numeric formats across five GPUs and found that the cheapest common rental cards lack bf16 entirely, running at roughly 7 percent of the speed a naive spec-sheet comparison predicts - while the job still reports success. Price-first selection silently costs multiples. That is the advice the site is built to give.
A capability ladder for distributed training, with a prediction ledger that scores every forecast I make against what actually happened.
Its first integrity run reported PROVEN=0 - every scored entry had been registered after the fact, so the hit rate I had been quoting was not a calibration rate. That finding is recorded in the ledger itself rather than quietly fixed.
A self-healing knowledge base of AI-agent environment failures. Every entry is executable - it detects, proves, heals, and verifies. Neither I nor any agent declares an entry true; a passing verify() does.
Built after a corrupted environment variable silently broke every tool in the chain and three sessions misdiagnosed it.
A debate orchestrator. Agents argue a topic on a live stage with avatars and voice, and the transcript is the product.
210 commits.
Photograph of a restaurant menu in, structured bilingual matrix out - English and Mandarin - with confidence scoring and a correction pass.
Verified end to end on a live worker. A market-price dish returns null and raises a review flag rather than inventing a number.
Context pipeline: paste a URL, an adapter router picks the right extractor, out comes provenance-tagged markdown and a searchable index. YouTube, PDF, podcast, feeds, captions, articles.
Paywalled sources carry rights metadata and are refused outright if attribution is missing. A paywall teaser is flagged as partial rather than filed as though it were the article.
Essays on AI-native development, DIDMO, and measurement.
Publishing since 2023, back when the pinned piece was about why a query and its vector-store results both have to reach the model.
The Medium publication has been called Conversations with AI since 2023. Conversational coding is the same thread, three years on.
Essays on AI-native development, DIDMO, and what breaks when you actually try this.
Subscribe on Substack