01
NovocanvoExport sales & AI automation
Built the prospect list from zero — 17 companies across 8 markets in the EU and US, 10+ cold outbound campaigns, every follow-up written and owned personally. That converted into real buyer conversations: a Danish manufacturer now sampling, a German buyer in active price negotiation. Then I automated the repetitive layer — a pipeline that re-scores every buyer as new data lands, triages inbound mail and drafts the reply, so effort keeps following the accounts most likely to convert.
Role Export Sales & Business Development
Scope FOB quotation across MOQ tiers (3,000–5,000 m; USD 1.30–2.40/m²) → order → shipment → payment collection
Export SalesCold OutreachAI Automation
02
F1 race engineerReal-time local AI · voice
A live race engineer for F1 25 that listens, thinks and talks back — entirely on my own machine, with no cloud AI involved. Whisper (large-v3-turbo) catches the radio and is bound to a custom phrasing table so it hears F1 vocabulary correctly rather than guessing. That feeds a two-part brain: a rules layer that answers instantly (where’s my rival, yellow flag, safety car) and a local model for the hard calls (undercut, strategy change, fuel and damage trade-offs). Output goes back out through TTS, so it actually speaks. It reports rivals on track and keeps damage, positions and fuel updated — and it was trained to understand F1 and refuse to invent what it doesn’t know.
Build Whisper large-v3-turbo → custom vocabulary → rules + local SLM → TTS
Training Qwen3.8-27B generated the training data and files, distilled down into a MiniCPM-2B run with reasoning off for latency
Runs on RTX 4090, fully local — no cloud AI at any point
Local LLMWhisperTTSF1
03
Personal AI assistantTwo personas, one local deployment
An always-on assistant I genuinely use every day, built as two switchable personas over a single memory and toolset. The Assistant side runs the day — file management, schedule checks, and the routine admin I would otherwise let slide. One click moves it to Soulmate: a warmer, conversational mode with its own tone, pacing and vocabulary. Both run on a locally deployed Qwen3.6 build, and the TTS voice was fine-tuned from real voice sources to the tone I prefer. Nothing leaves the machine.
Personas Assistant — files, scheduling, daily admin · Soulmate — conversational, different register and voice
Voice TTS fine-tuned from real voice sources to match a preferred tone
Runs on Qwen3.6, trained and deployed locally
Local LLMTTSAgentsMemory