I BUILD DIGITAL SYSTEMS THAT FEEL OBVIOUS AFTER THEY EXIST.

Turning an idea into working, owned software still starts with a blank editor. Non-technical founders and fast-moving builders either pay for a closed no-code tool they cannot leave, or wait weeks for scaffolding. SoftBox Studio closes that gap: describe the app in chat, watch an AI plan and build it, then refine it visually and export real, deployable Laravel + Vue code.

The core design decision was making the AI output structured and reversible, not free-form. Every turn returns a JSON action protocol (create_page, generate_crud, update_theme, update_database) applied transactionally, so partial failures cannot corrupt a project. A dedicated Context Engine compiles the whole application's memory into each system prompt, so the AI always reasons over the full app rather than a single page. Providers sit behind an AiProvider contract (multiple hosted LLM APIs, local OpenCode, and a keyless offline Fake), and every page is a JSON block tree rendered by the frontend - templates are never edited directly.

Phases 1 and 2 shipped: visual builder, version history with undoable snapshots, a three-tier component registry (30 primitives, 26 sections, 6 page templates), specialized agents (planner, architect, developer, designer, reviewer), a visual Database Designer, a Workflow Builder, and a Monaco-based code explorer that renders a real Laravel + Vue codebase with generated Pest tests. Phase 3 (multi-agent collaboration, Git integration, deploy, marketplace) is on the roadmap.

Queue-backed multi-turn auto-continuation with a rotating token to abort stale chains, solving PHP timeouts and runaway loops · Fully demoable with no API key via a deterministic offline provider · OpenCode local-AI server, Reverb WebSockets, project export, and a self-deploy admin panel

SMBs outgrow starter CRMs fast, then face a painful migration to an enterprise platform. SoftBox CRM was designed as a modular platform that scales from about 5 to about 500 users without a re-platform - every capability is a module you can enable or disable.

It is a modular monolith: each of 23 features ships as a self-contained module carrying its own routes, migrations, services, models, policies and tests, auto-discovered at boot by a ModuleServiceProvider and gated by config/modules.php. The frontend mirrors the same structure with lazy-loaded Vue pages and Pinia stores. The API is versioned (api/v1/), auth is Sanctum, and RBAC plus team_id multi-tenancy are built in. To avoid AI lock-in, all LLM calls route through an LLMProvider strategy (multiple hosted and local providers); per-module AI services (Sales, Emails, Deals, Analytics, Automation, Support) share one AiService, and prompts use schema-rich JSON to keep outputs reliable.

By the numbers: 102 models, 100 controllers, 50 services, 75 migrations, 108 Vue pages and 30 stores, backed by 54 PHPUnit test files across 15 feature-module suites plus core Unit/Feature suites on SQLite in-memory.

Automation engine with triggers, conditions and 10 action types · AI lead scoring, next-best-action and sentiment analysis · Omnichannel gateway abstraction (Twilio / mailgun) and an RFQ-to-procurement module · White-label platform module and mobile PWA

An Italian restaurateur runs the floor from paper, Excel, phone calls and sticky notes. Reservations live in one place, the menu in another, orders shouted to the kitchen, and the bill juggled at the counter. Tavolando replaces all of it with one real-time panel that mirrors the actual rhythm of an Italian service - from first booking to last bill.

Everything is real-time: 13 broadcast events fan out across Sala (floor map), Cucina (kitchen display) and Cameriere (waiter app) over private restaurant-scoped channels via Echo/Reverb. The product is Italian-first and treats money as integer cents throughout. The UI composes from a workspace design system package (57 primitives) whose tokens.json is the single source of truth, and the marketing landing is config-driven from the same config the app uses, so product truth is never duplicated.

70 tests and 80 migrations back 12 kitchen and floor pages, with real product screenshots rather than mockups. Pricing ships at 0 / 49 / 79 EUR, and the distinctive editorial landing shipped. A full rebrand from Tavolta to Tavolando swept roughly 150 files and 2,900 occurrences across four registered domains.

Money-integrity fixes: cancelled-item resurrection, correct 10% vs 22% VAT, discount overflow, and idempotent payments (409 + client key) · Realtime done safely: optimistic UI reconciled against the server with rollback and a polling/SSE fallback · Draggable, shaped floor plan built on pointer events with no canvas library · Accessibility-driven: a lime token at 6.62:1 contrast replaced a failing chart color

Italian legal professionals cannot track the roughly 424,000-document archive of the Corte Suprema di Cassazione by hand, yet a small fraction of rulings - especially those signalling nomofilachia (law-shaping) weight - matter far more than the rest. MassimAI monitors the archive, ranks each ruling by a 'could become law' score, and summarizes it.

Reliability was the guiding constraint. A transparent heuristic scorer runs synchronously on ingest, using weighted, human-readable signals (Sezioni Unite, referrals to a joint section, constitutional questions, questione di massima) so every ruling always has a defensible score. AI enrichment (massima and topics) runs as a best-effort async job that never blocks ingestion - if the provider fails, the row keeps its heuristic score. The crawler is deliberately polite: a 1,500 ms throttle, a contact-bearing user agent, exponential backoff with a kill switch, and a resumable page-cursor backfill. Before any text reaches the AI, PII is pseudonymized and rehydrated afterward; the mapping is never persisted.

115 backend test methods across 28 files and 137 frontend tests across 23 files run in CI, over 14 migrations, 26 routes and 8+ artisan commands. Search runs on MySQL FULLTEXT with a SQLite LIKE fallback for tests. Ingestion is scheduled (6-hourly fetch, weekly re-fetch/prune, daily 08:00 watchlist alerts).

Compliance-first: a documented legitimate-interest assessment, 365-day retention, a delist-vs-GDPR-erase split, a 30-day SLA, and a weekly delist sweep that strips text but keeps the score · Dual scoring behind one interface with deterministic fallback so ingestion never breaks · Watchlists and saved filters driving new high-band-ruling alerts