What I've built

The one fact that shapes everything: a Raspberry Pi 4B can't run a serious AI model. So Devalogam splits into a body and a brain. The body — web app, API and database — runs on the always-on Pi. The brain — a large language model — is kept remote and swappable (Ollama on a Mac mini by default, with Ollama Cloud, Claude or Gemini as fallbacks). The dashboard is fully usable with no brain at all: Kanban, Tasks, Calendar and Notes work entirely on the Pi, and the AI features simply light up when a brain is reachable.

The load-bearing component is Gub — an LLM router and meter. It health-checks providers in priority order, returns the first reachable one, fails over automatically, and meters every call for budgeting. Configuration, not code, picks the brain: switching from the Mac mini to Claude to a future bigger server is a one-line .env change, never a rewrite.

Capabilities grow like biology. An organ is a reusable function with a typed contract; an organism composes organs into a tool (Kanban, Calendar); a living organism is a running instance rendered as a card on the Land. One task organ feeds Kanban, Tasks and Calendar — implement once, compose three ways. A pantheon of ten AI agents (modeled on Hindu deities) governs the realm, all sharing one runtime and differing only by persona, data scope and tools.

Stack: a pnpm monorepo — a React + Vite PWA, a Fastify API, PostgreSQL 16 with pgvector — all behind a single Caddy reverse-proxy on one origin, packaged as Docker Compose. Security is Tailscale-only: nothing is exposed to the public internet, secrets never leave the Pi or reach the browser, and the whole system moves to a beefier host by copying a folder and a database dump.

Proves: end-to-end product ownership of a private AI platform — architecture, swappable-brain resilience, security posture and portability.

Everything feeds one number: today's score. Todos carry a point value you set and bank on completion; activities are time blocks tagged with a category that has a points-per-minute rate and a positive/negative flag — Deep Work might be +10/min, Doomscrolling −20/min. The day's net total maps to one of 13 configurable, color-coded tiers, from 💀 Catastrophic to ⭐ Legendary. Negative scoring is the point: most apps ignore the cost of wasted time; DORKprod puts it on the ledger so a day can't lie to you.

It's a strict separation of concerns: PHP 8 + MySQL handle data and auth only — a clean REST-style JSON API, no HTML rendering — while the entire single-page app (router, views, charts, drag-and-drop, text-to-speech, audio) is hand-written vanilla JavaScript with no framework and no build step. A two-way link ties the Today planner to the Kanban board: checking off a linked todo automatically moves its board task into the Done column, eliminating double-bookkeeping.

Security and operations are first-class: PDO prepared statements everywhere, hardened httponly/strict-mode sessions with ID regeneration, bcrypt password hashing, and every query scoped to the authenticated user's ID so users can only ever touch their own data. Access is deliberately invite-only — prospective users apply with a reason, and an admin approves or rejects from a built-in panel. Static assets are cache-busted by file modification time, so a plain upload to commodity shared hosting always serves fresh code.

Proves: full-stack product ownership on a conventional web stack — a novel mechanic, a secure normalized API, and a polished dependency-free front end.

Bus In ships understanding the vocabulary of growth — recurring revenue, retention, runway, efficiency ratios — instead of being a blank charting tool. On top of the curated metrics sits an AI analyst you own: natural-language Q&A, auto-generated insights and anomaly detection, an AI report builder, and AI-commented forecasting — all powered by your model (OpenAI, Anthropic, Gemini, or a fully local Ollama / LM Studio), so confidential board numbers are never sent through a vendor.

It's a client-side SPA: all metrics, computations and AI orchestration happen in the browser, with pluggable storage (memory, localStorage, or a self-hosted KV server you control) and real-data ingestion via CSV or HTTP/JSON. Built with TanStack Start, React 19, Recharts and TypeScript, it deploys as a static bundle to any host at near-zero cost.

Edi ma drives the actual FFmpeg engine compiled to WebAssembly — edits aren't toy approximations, they're the same operations professionals use, executed locally. Footage never leaves the device: no upload step, no render farm, no account. When the page is served cross-origin-isolated it automatically loads the multithreaded core for 2–4× faster encodes, and transparently falls back to the single-threaded core otherwise.

Every edit can be saved as a serializable .chop.json recipe — a human-readable, replayable, version-controllable description of the whole edit history. A real multi-track timeline (video / audio / text tracks with drag, zoom, snap, split and ripple-delete), color grading, overlays, chroma key and captions round it out. As an installable PWA it caches the ~31 MB FFmpeg core once and then runs fully offline.

Privacy is the architecture, not an afterthought: local-first by default, with optional self-hosted sync, and no server in the loop unless you deliberately add one. The toolkit is built for the hardest seconds of an urge — a panic button with breathing guidance, a crisis mode with a mandatory 60-second pause before a relapse can even be logged, and a guided post-mortem that turns a setback into data.

An AI urge coach and CBT-assisted reframing use your provider and your keys, calling the model directly from the browser. A one-click therapist pack exports a structured multi-week PDF of mood, craving, CBT and relapse data — turning private self-tracking into something genuinely useful in a clinical session. (Ene me is a self-help tool, not medical advice.)

Sel me treats a job search the way it actually is — a sales pipeline — and unifies applications, resume versions, cover letters, interview notes and a calendar of next steps into one keyboard-driven workspace. Its defining stance is data ownership: by default everything lives in your own browser; sync is an optional feature you point at a server you control. Your CV, salary expectations and recruiter conversations never pass through someone else's cloud.

Per-job tailoring is a first-class action: import a Word resume, maintain base versions, then create per-job overrides and export them back to .docx and PDF with the original layout preserved. Keyword matching highlights gaps against a job description, AES-256-GCM encrypts backups, and a ⌘K command palette drives the whole app. Built with TanStack Start, React 19, Zustand and TypeScript.