Things I’m building

Real investment memos take days. Most of that time is research most VCs do the same way every time — pulling Crunchbase data, sizing TAM, cross-referencing team backgrounds, sourcing expert commentary. So I built a coordinated pipeline of 33 specialized agents that turns a company name into an institutional-quality memo with full citation trails and firm-branded exports.

It is not a chatbot. Each agent produces a research-analyst artifact — a deck deconstruction, a Perplexity Sonar Pro source bundle, a section draft, a fact-check report — that lives as a file you can re-run, override, or audit. The orchestrator (LangGraph) coordinates them, manages state, and handles retries.

Bring your own Claude and Perplexity keys; deal data stays on your machine. Used by nine VC firms so far. The "draft a memo" → "polished firm-branded memo" loop is now ten to fifteen minutes for a first pass; one to two hours with an analyst on top.

Every AI session starts from zero. Vibe coding produces spaghetti. Prompt engineering is ad hoc. The thing missing is not a better model — it is externalized memory: documentation modular enough to fit in a context window and structured enough to be loaded selectively.

Context Vigilance is a deceptively simple framework: a context-v/ directory in every project, with four folders that pair into two cognitive modes. Specs and prompts handle planning — what you are building and the step-by-step path to get there. Blueprints and reminders handle reflection — codified patterns and the short, sharp corrections that keep AI from drifting.

When the work does not fit a known loop, two more folders appear — explorations for "I do not know where this is going yet," and issues for the half-day debugging journeys that "debug" does not honor. The kit, schema, and reference catalog are open source.

Twelve Astro sites in one workspace. Not a true monorepo — every site is its own git repository, independently deployable. Not a loose pile of repos either — co-located so patterns, brand conventions, and AI context can travel between them without coupling deployments.

The honest reckoning is part of the project. We tried the "one shared component package, all sites import it" approach. It collapsed under abstraction overhead the moment two sites needed different colors. So @knots/* packages stayed as workspace-local pattern references — code you copy and adapt, not import. Where sharing genuinely earned its keep was the markdown processing pipeline, which became the only published package: @lossless-group/lfm.

This site (mpstaton.com) lives in the workspace. So does FullStack VC, the Hypernova site, the Water Foundation, Cilantro, Dark Matter, and a handful of in-progress sites.

Every site I work on ends up wanting more from markdown — Obsidian callouts, citations that renumber themselves, bare YouTube URLs that become embedded players, link cards with build-time OpenGraph metadata. After writing the same render pipeline three times across different sites, I extracted it into a real package.

LFM is a polyglot remark/rehype pipeline. Authors keep authoring in whatever flavor they prefer (CommonMark, GFM, Obsidian, remark-directive); LFM normalizes the variations into one canonical AST. Syntax declares what you want; a trigger plugin parses it; one component renders it. Three steps, no MDX lock-in.

Published to GitHub Packages and JSR. The llms.txt endpoints, context-v document rendering, and demo Scroll Report on this site all run through it.

I write in Obsidian. So do most of the people I work with. Out of the box, Obsidian is excellent for thinking; for AI-assisted drafting, citations, image generation, and search-augmented enhancement, it needs help. Content Farm is what fills that gap.

Nine plugins that share a vault and complement each other: drafting helpers, citation tools, an image-generation bridge, OG-metadata enrichment, a context-loader for AI sessions. None of them try to be everything; together they form a content-creation surface that lets me move from a half-formed thought to a published piece without leaving the editor.

Open source, small footprint, no telemetry. If a plugin earns its place in the chorus, it stays; if not, it goes.

Most VCs I know are either "AI is a toy" or "AI is going to eat our jobs," and neither is a useful working stance. The middle path — actually practicing with the tools, in the open, with people doing the same thing — does not have many natural homes. So we made one.

FullStack VC is a learning community where venture capitalists deliberately practice using AI to give their firms superpowers. Monthly working sessions, recordings, working groups around specific use cases (sourcing, memo writing, portfolio support, LP reporting), and a coalition of people willing to share what is working and what is not.

It is not a course. There are no certificates. Show up, bring a real problem, leave with three things you can use Monday morning.