Bridging deep machine learning with high-performance production clients and backends.
15+ years architecting high-load web systems, distributed AI pipelines, and cross-platform desktop/mobile applications.
┌───────────────────────────────────────┐ ┌───────────────────────────────────────┐
│ AI / ML & INFERENCE │ │ CROSS-PLATFORM & CLIENT APPS │
│ PyTorch · NVFP4 · CUDA · TensorRT │ │ Flutter & Dart (Desktop, Mobile, Web│
│ LTX-Video 2.5 · Multimodal LLMs │───│ 100+ Production Builds · TDLib / FFI│
│ Quantization · Low-Latency STT/TTS │ │ Real-Time Copilots · Clean Arch │
└───────────────────────────────────────┘ └───────────────────────────────────────┘
│ │
└─────────────────────┬─────────────────────┘
▼
┌───────────────────────────────────────────────────────────────────────────┐
│ DISTRIBUTED BACKENDS & MLOPS │
│ FastAPI · Docker · Multi-GPU Clusters (SaladCloud / RunPod) │
│ Resident VRAM Orchestration · High-Concurrency WebSockets │
└───────────────────────────────────────────────────────────────────────────┘
- 🚀 TellMeMore — Real-time AI interview co-pilot and floating teleprompter for macOS built with Flutter & Dart. Features dual-stream live speech-to-text (Deepgram Nova-3), zero-latency prompt synthesis, screen-share stealth protection (
NSWindow.SharingType.none), multi-LLM orchestration (Gemini, Claude, OpenAI), and unit test coverage. - 💬 TellMeMore TG — Cross-platform Telegram client powered by Flutter + TDLib with an embedded multi-agent AI copilot for contextual dialogue profiling, manipulation detection, and strategic negotiation assistance.
- ⚡ ltx25-fast-docker — Production Docker container & FastAPI inference microservice for LTX-Video 2.5 (22B Distilled). Delivers ×2.1 faster wall-time inference on 24 GB VRAM (RTX 4090 / 5090) via Blackwell-native NVFP4 quantization, Gemma 4 Heretic TE, reduced 6+2 sigmas, and resident memory orchestration.
- 🌐 tellmemore-chrome — Universal real-time AI copilot & strategic negotiation assistant WebExtension for browser-based messengers (TypeScript, DOM streaming).
- 🎨 qwen_upscaler & ltx25_lite_gen — Production generative studios: Qwen 2.1 Turbo (NVFP4 ~1.7s T2I), Qwen3-VL 8B multimodal editing, and SeedVR2 4K video/photo super-resolution.
- Production Fleet: 100+ production builds and releases across desktop (macOS, Windows) and mobile (iOS, Android).
- Native & Hardware Interop: Deep integration with C/C++ libraries via
dart:ffi, TDLib bindings, native macOS window APIs (macos_window_utils, accessibility APIs, screen capture exclusion), and dual-stream audio capture. - Architecture & State: Clean layered architecture (BLoC, Riverpod, Provider), reactive streaming, WebSocket/SSE state sync, offline SQLite/Isar caches, and thorough unit testing.
- Inference Optimization: Distilled diffusion and video transformer tuning. Optimized LTX-Video 2.5 inference by ×2.1 on RTX 4090/5090 hardware (NVFP4 mixed quantization, Gemma 4 Heretic TE, reduced 6+2 sigmas, resident VRAM scheduling).
- Headless Production Pipelines: Decoupled ComfyUI server clusters with automated REST APIs, WebSocket job trackers, and dynamic image/video resolution scaling.
- Diffusion & Upscaling: Production pipelines using Flux.1 Schnell/Dev, Qwen Turbo, and SeedVR2 4K temporal video/photo upscalers.
- Custom Models: Fine-tuning via custom LoRA architectures, ControlNet conditioning, IP-Adapter styling, and dataset curation.
- Containerized Clusters: Multi-GPU orchestration across SaladCloud, RunPod, and dedicated bare-metal nodes.
- Polyglot Microservices: High-performance backends in Python (FastAPI, PyTorch), Node.js / TypeScript, and Go / C++.
- Zero-Downtime Releases: Automated CI/CD pipelines, containerized inference runners, and health check monitoring.
- Hardware Efficiency: Minimizing VRAM fragmentation, cold-start latency reduction, and multi-tenant resource sharing.
| Domain | Technologies |
|---|---|
| Client & Mobile | Flutter, Dart, Android SDK, iOS (Swift / Obj-C), macOS APIs, FFI / C++, BLoC, Riverpod |
| AI / ML & LLMs | PyTorch, Hugging Face, CUDA, TensorRT, NVFP4, LTX-Video, Flux, Deepgram STT, Gemini, Claude, OpenAI |
| Backend & APIs | Python, FastAPI, Node.js, TypeScript, Go, C++, WebSockets, Server-Sent Events, REST |
| Cloud & DevOps | Docker, Docker Compose, SaladCloud Multi-GPU, RunPod, Linux, Nginx, Redis, PostgreSQL |
- 💬 Telegram: @s3rf__im
- ✉️ Email: s3rf.dev@gmail.com
- 🌐 Portfolio: s3rfdev.github.io
- 💼 Available for: RLHF / LLM Evaluation & Alignment, AI/ML Consulting, Cross-Platform Architecture, and High-Performance Engineering.