AI Production — Case Library

Generative 3D Pipeline — 10M+ Item Catalogue Production

Catalogue of unique 3D game items (GLB models with 2D renders) for a commerce platform. Second-iteration pipeline cut production cost from ~€2.8M to €400K: concept generation through image-to-3D, procedural multiplication, automated QA, and headless rendering. Visual consistency enforced across genre sets — silhouette, materials, palette, detail level.

Stack: FLUX + custom LoRA · Rodin / Meshy / Tripo · Blender Geometry Nodes · DINOv3 / OCR / VLM validation · Cycles + OptiX headless · Prefect 3 · PostgreSQL. Dual-track sourcing strategy (commercial enterprise APIs vs self-hosted open models) with licensing and IP clearance as first-class architecture constraints.

24/7 AI Community Host Production

Multi-agent system operating live communities around the clock across Discord, Telegram, and WhatsApp — moderation, engagement, member support — replacing a manual community-management function.

Stack: Agent orchestration · RAG · platform API integrations.

AI Front-Desk for HoReCa Group Production

Agent system automating ~250 monthly bookings and food orders end-to-end for a 4-location restaurant group — intake, confirmation, changes, escalation to staff only on edge cases.

Stack: Multi-agent orchestration · business-tool integration · messaging channels.

ENIGMA — B2B OSINT Early-Warning Platform Architecture & Spec

Investigative-intelligence platform: natural-language monitoring contracts («notify me when [entity] [event]») over public data sources — SEC EDGAR, GDELT, X, Reddit and others. Precision-over-completeness alerting: one alert, as early as possible, with an evidence trail. 18-agent architecture in three layers (ingestion / analysis / delivery), dual-LLM validation, ~90% deterministic decisioning with LLMs reserved for ambiguous cases. Full technical specification delivered for implementation.

Positioning: between free alerts (Google Alerts) and enterprise suites (Brandwatch, Meltwater) — for investigative journalists, analysts, and investors.

WorkMesh — AI Team Assembly Architecture

AI-driven team assembly from a 2,000-contractor pool with a 2-day onboarding and offboarding cycle — replacing a weeks-long staffing process with LLM-powered matching and orchestration.

Stack: LLM orchestration layer · contractor-pool data model.

Real-Time Speech Translation for Live Calls Working tool

Browser-based bidirectional live translation (EN/PT/RU) for client calls in Teams / Meet / Zoom: dual-channel streaming STT, LLM semantic gating before phrase commit, streaming translation, low-latency TTS routed back into the call as a virtual microphone.

Stack: Deepgram streaming STT · Claude (gating + translation, SSE) · ElevenLabs TTS · AudioWorklet / virtual audio routing — zero server-side.

In-House Stable Diffusion Pipeline (X-FLOW) Production

Custom-trained in-house image-generation pipeline for game art production — eliminated €700K in external art costs over 12 months and cut concept-iteration time by 40%.

Stack: Stable Diffusion · custom training · LoRA · production workflow integration.