AIdemica

Gallery

About

WHAT IT ISAIdemica is a live production platform that turns a university assignment brief into a near-finished draft package: a Word document, a slide deck, a defense speech and A1 engineering drawings. The user does the final pass; the system does the other 99%.WHY IT IS NOT A CHATGPT WRAPPERThe core is a custom MCP server driving a three-wave parallel agent swarm, wrapped in an enforcement layer that physically prevents low-quality output from shipping.Wave 1 - one sub-agent per chapter: research and calculations, returns structured JSONWave 2 - coordinator de-duplicates figures across chapters, one agent per figureWave 3 - parallel section writersTail - coordinator writes intro, conclusions and bibliographyEvery figure must carry a justification AND evidence: a quoted page from the brief, a sample id, or the filename of computed results. Fewer figures than planned is a hard failure, not a warning.THE ENFORCEMENT LAYER (the real IP)- 12 automated validators: section mapping, citation coverage, humanization diff, block-formula preservation, broken figures, figure references, numeric consistency, work-type matrix, citation survival, physical-magnitude sanity, figure index, subsection policy- Stage gates aggregate ALL issues into ONE rejection instead of round-tripping the model seven times- SHA-256 tamper protection on stage flags: the model cannot raise its own page limit to sneak an oversized document through- Builder-marker checks: the model cannot substitute a hand-rolled script for the official document builderTWO IN-HOUSE ENGINES- Presentation engine: constraint solver, 18 layouts, 12 themes, WCAG-validated colour, seed-based variation- CAD engine: text to DXF/DWG, 18 vectoriser iterations, GOST 2.104 title block generated in code- Humanizer: 150+ patterns across 6 languages, wired into the validator so formulas and variable names cannot drift during rewritingSECURITY- Tenant isolation enforced at two layers (permission deny rules plus a PreToolUse hook): an agent cannot read another order- White-label: the assistant never discloses the underlying model, tools, paths or infrastructure- Prompt-injection gate; uploaded code is read but never executed- 154 security audit findings closed across 8 phases, GDPR endpoints, SentryCOMMERCIAL LAYER- 5 payment integrations written and tested: Stripe, Paddle, CryptoBot (USDT/TON), LiqPay, Fondy. Payment triggers the AI run automatically- Pricing engine: 24 work types, 100+ disciplines with individual coefficients, 15-point urgency curve, uniqueness premium, per-sheet drawing pricing- Promo codes, reviews with promo issuance, admin panel with refunds- 4 interface languages, multi-currency with FX table- Telegram support bot with email fallback, 6 templates x 4 languagesINFRASTRUCTUREDocker plus Caddy with automatic HTTPS, SQLite in WAL mode (Postgres-ready), nightly backups with 30-day retention, 44 restore points on disk. Running healthy on a dedicated server. Domain included, registered through May 2027.TRACTION - stated honestly- ~2,960 unique visitors per month, 83 to 248 per day, fully organic, zero ad spend. Largest share of requests from the United States, then the Netherlands, France, China and Canada- 33 completed orders on disk, 867 MB of delivered work- ~55,000 lines of core source with about 23 TODO/FIXME markers in totalWHAT THIS IS NOTNo revenue. The payment integrations are written and tested but no live merchant keys are connected: the buyer plugs in their own accounts, which is how it should be. Order flow stopped in June when I moved to other projects. The presentation engine has 4 of 15 variation knobs not yet wired (documented, roughly 3-5 days of work).I am selling a working system with real traffic, not a running business.WHY I AM SELLINGI built this solo and moved on to other projects that take all my time. It needs an owner who will connect a merchant account and work the traffic that is already arriving.