PortfolioV2

BuilderOperated end-to-end

Est 2026

Matt
Torcasso

I design, build, and operate production software end-to-end. A builder who ships: idea to working, deployed product — close enough to the architecture to debug it, extend it, and keep it running in production.

matthewtorcasso@gmail.com entangled.chat

Selected work

Trainer App

01

A voice-first SaaS platform for personal trainers: 1,287-exercise library, mesocycle programming, and a 25-tool AI agent layer. Live and in daily use.

View project →

April

02

A personal AI operations platform: one agent brain on two surfaces (Telegram and voice), ~100 tools, ten-plus systemd services on a home Linux server.

View project →

Governor

Governor

03

A meta-agent over three walled-off trading strategies, live on a real funded account with an HMAC-signed human approval gate on every order.

View project →

Entangled

04

An anonymous AI confidant, live at entangled.chat. Four-word identities, a burnable memory page, and a three-layer safety wall that fails safe.

View project →

Ledger

Ledger

05

A self-hosted personal finance PWA: file-import ingestion, recurring-charge intelligence, and a strictly read-only MCP server over one SQLite file.

View project →

Jerry

06

An autonomous job-search agent that hunts nightly, drafts through a lint-gated document pipeline, and has honesty enforced by code — it can only propose.

View project →

Also running

Usage dashboard

Tracks exactly what my AI tooling costs, per project and per model, from Claude Code's local logs. Tokens are real money, and I run this operation like it.

Home infrastructure

Everything above runs on hardware I administer: a Linux mini-PC with systemd services and timers, Tailscale networking, Cloudflare tunnels, and health checks that fail loudly.

The common thread

Every system here ships with rails I designed: signed approvals, confirm-gated writes, hard budget caps. Building automation is the easy part — making it trustworthy enough to run against real money, real clients, and real data is the work.

How I work

I'm not the coder, and I'm upfront about that. What I bring is translation: I can take a technical topic, understand the infrastructure and the logic underneath it, use the tools hands-on, and see clearly where AI genuinely helps and where a simpler solution wins.

I'm an AI realist and a believer, and above all a connector, the bridge between the technical minds who build the tech and the people who actually need it. Closing that gap is the work I'm best at.

Selected work · 01 / 06

Trainer App

In production · built to sell

A SaaS platform for personal trainers, live and in daily use by a working trainer and his clients. The wedge is speed: program a client's next training block by voice or through an AI agent instead of clicking through forms. One codebase serves three perspectives — the trainer's workstation, a login-free client portal, and tiered share links with PDF export.

1,287exercises in the library

25AI agent tools (MCP)

3surfaces, one codebase

2hself-destructing demo tenant

Trainer App client analytics hub: lifetime weight moved, weekly session metrics, and an automated coach's read

Mesocycle template library: reusable training blocks filterable by resistance, cardio, and hybrid

Week calendar with planned and scheduled client sessions (names anonymized)

Client analytics hub, mesocycle template library, and the week calendar. Client names anonymized.

What it does

Voice-first floor logging: dictation runs through a deterministic parser with a Claude Haiku fallback, so sets get logged mid-session without breaking stride
Mesocycle builder with abstract movement slots resolved per client, supersets, and a shelf of reusable starter programmes
Macrocycle layer: objectives, three assessment types (FMS, body composition, cardiovascular), and an adaptive nutrition engine built on Katch-McArdle BMR
Client analytics hub with a custom metric engine and reusable chart kit — per-client, per-meso, plus business funnels for the trainer
Client portal & share links: magic-link access with no login to manage, tiered plan sharing, PDF export, Apple Watch ingestion

Under the hood

25 MCP tools expose the whole app to an AI agent: scheduling, programming, plan surgery, notes, cardio logging, the full mesocycle lifecycle
Every write is confirm-gated: two-phase, idempotent, and nothing deletes — the agent proposes, the trainer approves
Phone-first PWA; a feature is not done until it works on a real device on the gym floor
Invite-gated signup with per-IP rate limits, and a demo tenant that mints on request and destroys itself after two hours
Next.js 16 App Router, TypeScript, Tailwind v4, Supabase (Postgres + magic-link auth), Vercel, Vitest

Next.js 16TypeScriptSupabase PWAMCP / AI agentsB2B SaaS

Live demo available on request

Selected work · 02 / 06

April

Live · daily use

My personal AI operations platform — one agent brain living on two surfaces that share the same memory: Telegram for text, and a voice server driven by an iOS Shortcut over Tailscale that works hands-free in CarPlay. It runs my calendar, briefings, nutrition, reading, job search, and the trading approval loop, from a Linux mini-PC in my house. It is how I found out this whole approach holds up over months of real use.

~100tool definitions, 27 modules

71Python modules

10+systemd services & timers

2surfaces, one shared brain

April HQ in demo mode: a morning briefing over a scenic backdrop, with schedule, training, money, and messages at a glance

April's money panel in demo mode: portfolio value, gains, holdings, and a chat with the money-desk agent

April HQ, the web dashboard — hub and money pages, shown in demo mode.

What it does

Morning and evening briefings, email triage, capture processing, market watch, and weekly reflection — each its own scheduled systemd job
Hands-free voice with Whisper speech-to-text and Kokoro text-to-speech, reachable from the car over Tailscale
April HQ, a themed web dashboard: hub, money, jobs, nutrition, health, and portal pages with real-time status
One-tap approve/reject controls for the trading system's pending orders
Domains for money (Warren), job search (Jerry), nutrition, calendar, and reading — all the same brain, different desks

Under the hood

Memory bank and playbook persist across sessions; conversation state resumes a single Claude Code CLI session
Every scheduled job stamps a heartbeat; a health-check turns each one green or red, so silence is never mistaken for success
Runs as systemd services designed to survive reboots and fail loudly rather than silently
MCP integration with my Obsidian vault for notes and records
Python, Claude Code CLI, Whisper + Kokoro, Tailscale, systemd

AI agentsPythonLinux / systemd VoiceTailscaleAutomation

Selected work · 03 / 06

Governor

Live · real account

A meta-agent over a strategy lab with chinese walls: three deterministic paper strategies, each testing one of my market theses in its own walled-off account with its own budget and scoreboard. The Governor observes them daily, surfaces only the trades the theories rate best, and proposes them against my real, funded brokerage account — live since July 14, 2026. It proposes; I sign.

3walled-off strategies

Hard railsdrawdown halts & order caps

HMACsigned order approvals

Jul '26live on a real account

The approval gate

The Governor only proposes. Every order waits in a pending queue until I approve or reject it — a tap in April HQ, a text to April, or a CLI command
Every approval is HMAC-signed. The bot quarantines any approved order whose signature doesn't verify; order ids must match in full
Approval text is parsed by plain code, never by the model — the AI cannot approve its own trades by generating the right words
Hard rails get the last word even after approval: order caps per day, no sub-$1 stocks, a 25% drawdown-from-peak halt on all buying, per-strategy exposure limits

The strategy lab

Three theses, three walls: each strategy tests one market thesis in its own paper account with its own budget and scoreboard — no strategy can touch another's money or hide inside another's results
Benchmarked honestly: everything is scored against simple SPY buy-and-hold
It grades itself: weekly, the Governor scores its own past calls into a learning file that feeds its future prompts
Judges each paper strategy as working, losing, or unproven; proposes parameter fixes for losers, never auto-applies them
Python, brokerage API, public-filings signal fetchers, systemd

Trading systemsRisk controlsHuman-in-the-loop Pythonsystemd

Selected work · 04 / 06

Entangled

Live · entangled.chat

An anonymous AI confidant — a safe first place to just let it out, with peer matching between real people as the destination. No name, no account, no one watching, free for the people using it today. A mission project, live since July 23, 2026, served from my own hardware behind a Cloudflare tunnel.

4 wordsare the whole identity

3 layersin the safety wall

0emails or passwords stored

Hard capon daily AI spend

Entangled landing page: Everyone needs a place to just let it out. No name, no account, no one watching, free forever

The landing page and a card from the brand system — both generated in-project.

Privacy by design

Pseudonymous from the first commit: four common words are the entire account — no email, no password, nothing to breach
A visible memory page: the AI's memory of you is shown to you in plain language, and every line is burnable — deleted lines are fed back as "never re-add"
Retain then purge: raw conversation turns are deleted once folded into the memory page
Safety logs record level and category, never content; two separately-salted hash logs cannot be joined

Safety & operations

A three-layer safety wall: a regex floor, a Haiku classifier, and a Sonnet escalation that can raise the alert level but structurally cannot lower it. It fails safe
Abuse floor: per-user and per-IP rate limits, capped session minting, and a daily token-budget kill-switch
Built to grow into a self-sustaining business without ever charging the person in crisis
Node.js / Express, SSE streaming, SQLite, phone-first PWA in dark and light

Node.jsPrivacy / safetyAnthropic API SQLiteCloudflarePWA

Visit entangled.chat →

Selected work · 05 / 06

Ledger

Live · self-hosted

A self-hosted personal finance tracker — my Quicken replacement. File-import first, better recurring-charge detection than the product it replaced, an investment and net-worth view, and a phone-installable PWA. All data lives in one SQLite file on my own hardware: no cloud, no accounts, no fees, and no screenshots here on purpose — it holds my real finances.

9read-only MCP tools

1 fileof SQLite holds everything

3import formats: CSV / QFX / OFX

0cloud dependencies

Recurring intelligence

Merchant normalization groups charges by cadence and classifies them: subscription, bill, income, or transfer
Price-creep detection flags any recurring charge running more than 5% above its trailing median
Zombie detection surfaces subscriptions that stopped billing — the ones you forgot you had
Internal transfers auto-pair, so money moving between my own accounts never reads as spending

Security posture

The AI can look but never touch: the MCP server opens the database in read-only mode — nine query tools, zero write tools
Tailscale-only access: the web app binds to the Tailscale interface, not the open internet; tailnet membership is the access control
Nightly systemd sync pulls prices (CoinGecko for crypto, Yahoo for equities)
Python, FastAPI, SQLite in integer cents, vanilla-JS PWA with hand-rolled SVG charts — no build step

PythonFastAPISQLite PWAMCPTailscale

Selected work · 06 / 06

Jerry

Running nightly

My autonomous job-search agent, named for Jerry Maguire: one client, full attention, uncomfortable honesty. Every night it hunts, scores what it finds against verdicts I've given it, and drafts the shortlist I wake up to. If a well-matched application led you here — that was Jerry. The hard line: Jerry proposes, and only I submit.

Nightlyhunts run while I sleep

Directreads employers' own systems

3 stagesautomated document pipeline

1 builddesigned & run end-to-end

Jerry's desk in April HQ: weekly progress meters, search analytics, and a slate of proposed roles awaiting a verdict, each with Jerry's fit score, his honest read, and tap-to-decide controls

Jerry's desk in April HQ — the nightly slate awaiting verdicts. Shown with fictional demo listings.

The hunt

Reads employers' own hiring systems via their APIs and headless Chromium — so every posting on the slate is confirmed open at the source, not scraped from an aggregator
Every proposal comes with an honest read: Jerry's fit score, why it matched, and any qualification flags — surfaced, never hidden
Verdicts I tap on each proposal (yes / no / maybe, and why) feed a learning loop that sharpens the next night's slate
Progress meters keep the week honest, and the whole desk lives inside April HQ with one-tap decisions from my phone

Honesty, enforced by code

Propose-only by construction: agents can only mark an application "Drafted" — only my word sets "Submitted"
Three-stage documents: Claude Opus drafts structured JSON, a linter enforces mechanical voice rules and loops until clean, then a template renders the final .docx
The linter enforces anchor floors on every resume — no invented experience gets through
Python inside April, Claude Opus, Node lint/template pipeline, systemd timers, Obsidian as the system of record

AI agentsPythonHeadless Chromium Claude OpussystemdObsidian