Golden-hour view from a wildflower hilltop over a curving beach — a couple and their dog walk the shore, a small plane rests on the bluff, and three crosses stand on the far headland
Portfolio · Lucas Cashwell

Proof

Everything I've actually built — led by DATproof, Skillproof, and Modelproof.

See the proof Say hello
Projects

Things I've built

Three builds lead the work — DATproof, Skillproof, and Modelproof. Around them: a developer tool that ships on npm, a full-stack app I built to learn on, and a few earlier finance prototypes.

Live · rebuilt daily

DATproof

DATproof is an adoption accelerator for digital asset treasuries — an information hub for corporate Bitcoin purchase data.

  • The record. Corporate Bitcoin purchases, and the raises that fund them, straight from company disclosures.
  • Always current. The site rebuilds itself from the filings every day.
  • The credit side too. STRC and SATA — digital credit — tracked alongside the buys.
PythonSEC filingsGitHub ActionsStatic site
Live site ↗ GitHub ↗
$56.2B
raised at the market ·
184 filed sales
848,986
BTC bought · 180 filed buys
as of Aug 2026
Live · self-refreshing

Skillproof

Give your agents the skills they need. A catalog of community skills, plus a Claude Code skill that installs the right one into the setup you already have.

  • A plan you approve. It proposes what fits; you approve each part of the plan before anything changes.
  • It fits your setup. It reads what you already have and integrates — it doesn't pile on.
  • It maintains itself. A GitHub Action refreshes the catalog every day.
Claude Code skillMCP serverVanilla JSGitHub Actions
Live site ↗ GitHub ↗
claude — skillproof
> my sites all look the same
● Skillproof: found 3 · best fit impeccable
● one clash: your CLAUDE.md bans styling advice
  Install + soften that line? › 1. Yes
✓ installed · triggers verified · undo saved
Live · independent

Modelproof

Which AI model should you actually use? A calm, independent answer, matched to your work and your budget.

  • An advisor, not a leaderboard. Neutral, cost-first, and built around the tools you already pay for.
  • Sourced. Prices come from vendor pages; a blank means not verified.
  • Independent. No lab sponsors it; picks are on merit and cost only.
Vanilla JSClient-side52 models · 25 vendorsStatic site
Live site ↗ GitHub ↗
The buy zone · price vs capability52 models
← cheaperpricier →
Top-left = cheap and capable. Pricing verified · benchmarks flagged · blank = not sourced.

rls-guard

📦 Published on npm

An MCP server you point at your own Supabase project. It tells you, in plain English, which tables the public key can read — then hands you the exact SQL to lock each one down. No dashboard, no sign-up; your AI assistant just runs it.

  • It ships where agents find it. Agents discover it in MCP registries and suggest installing it — the new app store for dev tools.
  • Built from real work. It knows an accidentally-public table from a deliberately-public one — straight out of the Supabase RLS I wrote by hand for Prompt Emporium.
TypeScriptMCP SDKnpmPostgRESTNode.js
GitHub ↗ npm ↗
Supabase RLS scan2 exposed
profiles🔒 protected
orders⚠ public read
api_keys⚠ public read
audit_log🔒 protected
-- fix, ready to paste
alter table orders enable row level security;

Prompt Emporium

📦 Archived

The project I built to teach myself full-stack. A web app for storing, versioning, and AI-improving prompts and Claude skills — auth, payments, a community library, the whole thing. Tabled now — but it's where most of my hard-won lessons came from.

  • My first full-stack build. Next.js, Supabase (auth + Postgres + RLS), Stripe, Vercel — all wired up from scratch.
  • Where I learned to direct AI agents. Built over a lot of Claude Code sessions; this is where I figured out how to get consistent, production-quality output.
  • And where deployment finally clicked. Vercel config, Supabase environments, OAuth, webhooks — most of the hard config lessons came from here.
Next.js 15TypeScriptSupabaseAnthropic APIStripeVercel
Prompt versionsv1 → vN
Each prompt versioned + AI-improved over time

BTC Intelligence Suite

◆ Prototypes

Three Streamlit apps I built to dig into bitcoin-treasury intelligence — MSTR's capital structure, corporate treasury modeling, and adoption tracking. Real public 8-K data where it counts.

  • MSTR Bond Decoder. Decomposes MicroStrategy's convertible-note stack from real 8-K filings — NAV, BTC yield, conversion premiums.
  • Treasury Optimizer. Models optimal BTC acquisition for any corporate treasury; stress-tests against historical drawdowns.
  • Adoption Tracker. Scores S&P 500 companies on treasury-adoption readiness and drafts Strive-style activist letters.
PythonStreamlitPandasPlotlyPublic 8-K data
MSTR bond decoderprototype
BTC held
576,230
BTC NAV
$55.9B
Convert tranches
23
NAV premium
+187%
Parsed from real MSTR 8-K filings · illustrative
About

A little about me

I work in professional services, specializing in finance and SOX. I got tired of watching the gap between what AI could do and what people were actually doing with it. So I started building on nights and weekends to see what I could actually ship.

These projects are how I learn. Some are polished, some are rough — all of them taught me something real. That's what this site is: the actual work.
Contact

Say hello

Always up for talking AI, building, or whatever you're working on.

Email me LinkedIn
lucascashwell3-ai github.com