Still doing it all
by hand?

Rebuilt as a modern app — public site + private client area, in three languages (IT/EN/ES) — from a site that barely worked.

Every project gets its own room — versioned audio, comments pinned to the exact second, downloads reserved for paying clients.

Delivered hardened — automatic checks on every change, secured, anti-spam contact form; two access leaks found and closed.

You learn by speaking — a real conversation, out loud, in English or Spanish — from minute one. Almost no text, by design.

It remembers you — every call is mined for the mistakes you actually made; spaced repetition resurfaces the weak ones and picks tomorrow’s scenario.

Practice built from your errors — writing, translation and vocabulary drills are generated from the mistakes you actually made — graded by an LLM, and only the confident calls feed your memory.

Four agents, not one prompt — absolute beginners get an agent of their own — slower voice, listen-and-repeat, no corrections.

Describe a mood — in plain language or tap a 12-emotion wheel — a Claude agent resolves it and Lyra builds the playlist as a trajectory through that space, not a flat list.

Chosen by the lyrics — sentence-transformer embeddings place each track on a valence × energy taxonomy; the exact cited line that matched your mood is surfaced via Musixmatch.

Steer live — more like this / change the mood / raise the energy — reshaping the queue without cutting the current track; a 3D compass turns to your mood and traces the path.

In-page checkout — waveform previews, a multi-license cart (MP3 / WAV / stems / exclusive) and Stripe without leaving the page.

Delivers itself — every purchase sends private download links, valid only for the buyer.

Running costs cut to the bone — audio flies from Cloudflare straight to the browser, no server in between.

Answers from real data — ask in plain language about clients, orders or calls — the agent reads the CRM, the ERP and the documents and cites its sources.

Math done by code, not the AI — the model only picks where to look and how to explain; the numbers are always exact.

Top 10 at Cursor × Yellow Tech — qualified for the Italian National Hackathon League.

Handles the call on its own — recognises the student, gives course info, books, cancels and reschedules — and hands to a human when needed.

Italian and Spanish, no talk-over — answers instantly; confirmations and reminders go out by SMS.

Tested before going live — 9/9 real scenarios passed, with a panel tracking the time and cost of every call.

Collects the booking in-browser — chats with the user to gather the details before calling.

Calls the restaurant itself — autonomously dials via Twilio and handles the full phone conversation end-to-end.

Remembers you — persistent cross-session memory that survives ElevenLabs’ stateless calls — recognises returning customers and proposes “the usual”.

Admin dashboard — session analytics and revenue tracking.

Headless CLI — point it at a producer’s drop folder and it publishes each beat to Beat Store through the store’s authenticated API.

AI metadata + SEO — enrichment via Claude API; idempotent, resumable upload pipeline.

You confirm before publish — it reviews every drop in the terminal and asks first.

Interpolate in latent space — upload up to 4 audio files and blend their RAVE latent encodings via a 2D board.

Barycentric weighting — click position sets the blend across all inputs simultaneously.

FastAPI + TorchScript — backend on HF Spaces, React frontend on Vercel.

Analysis → generation — a three-client pipeline: 3D mood visualization → emotion mapping → prompt-free AI generation, looping in real time.

Sacred geometry as interface — Metatron’s Cube and Platonic solids map emotions to audio features in an interactive 3D client.

Prompt-free generation — via Suno API; comparative benchmark of generative models (Suno, RAVE, MusicGen, Jukebox).

First WDF on the H9000 — a Wave Digital Filter algorithm modeling linear and nonlinear circuits directly on the hardware via VSig3.

Nonlinearities via CPWL — a Canonical PieceWise-Linear representation; validated end-to-end with a diode-clipper circuit.

No prior art — built within VSig3’s low-level limits with no examples in the literature.

Automates invoicing — via Fiscozen, and sends personalized client emails.

Syncs & books — Google Calendar sessions → Notion; auto-books on third-party platforms.

Runs daily — across invoicing, email dispatch, session logging and booking.

Two agents, live escalation — Aria handles passengers on the front line and escalates complex cases to Marco, a supervisor agent.

Multilingual — conversational support (IT / EN / ES / FR) on ElevenLabs Conversational AI.

Hackathon build — made at Yellow Tech × ElevenLabs.

Genre from MFCCs — classifies tracks with librosa features, trained on the GTZAN dataset.

NN vs CNN vs RNN-LSTM — compared with accuracy/loss evaluation; scalable to other datasets and models.

Act, see the impact — users mimic everyday actions and watch their CO₂ and climate impact in real time on a responsive 3D globe.

Gesture → visuals → sound — Python gesture recognition + TouchDesigner real-time 3D + SuperCollider generative soundscapes.