Alexey P.

I fine-tune and serve open models on your own GPUs: Whisper and open LLMs (Qwen, Llama), vLLM inference, quantization, evaluation. Production AI agents and call analytics on top. Everything measured, nothing taken on faith. PROOF, NOT PROMISES • Lead engineer of a call transcription and AI analytics platform for a cloud PBX company: 5,000+ hours of calls a day, millions of calls processed, fully on-premise on our own GPU servers. Designed, built and still run it end to end. • Whisper large-v3 fine-tuned for a low-resource language: data collection, cleaning, self-training on unlabeled calls, ~380 h of training data. In production since August 2026, on par with a commercial ASR service specialized in that language. • Self-hosted LLM analytics on vLLM (Qwen 27B): NVFP4 quantization on Blackwell GPUs, speculative decoding (multi-token prediction), prefix caching, LLM and Whisper sharing the same GPUs. Call summaries, KPIs and follow-up flags with strict JSON output and every conclusion backed by a quote from the transcript. • Own transcription engine on CTranslate2: batched Whisper inference, long-call continuation, separate stereo channels, fixed VAD timestamps, concurrent transcription on shared model weights. • 250+ ASR evaluation runs: Whisper family (tiny to large-v3, turbo, distil), my own fine-tunes, NVIDIA Parakeet and Canary, Qwen3-Omni, Qwen3-ASR, Gemini and commercial APIs, on real calls in five languages. Own WER tooling with word-level diffs. Also trained Parakeet, Canary and Nemotron streaming models. • Upwork projects in the last year: WhisperX diarization and hallucination fixes; transcription of a 32 TB podcast/YouTube archive; a benchmark of 9 commercial speech-to-text APIs; an automated transcription portal; an AI sales agent on LangGraph. WHAT I DO 1. Fine-tune open models for your language, accent or domain Whisper, Parakeet, Canary; open LLMs with LoRA/QLoRA or full fine-tuning. Dataset building, cleaning, forced alignment, train/eval splits, honest before/after metrics (WER, task accuracy). 2. Serve open LLMs on your hardware with vLLM Model selection, quantization (NVFP4, AWQ, GPTQ), speculative decoding, prefix caching, memory tuning, GPU sizing, throughput and cost-per-million-tokens benchmarks. Private by design: your data never leaves your servers. 3. Make transcription faster, cheaper and correct Batching, CTranslate2, GPU utilization, hallucination and repetition-loop fixes, diarization, timestamps, language routing, hold-music and voicemail handling. 4. Build production AI agents and LLM applications LangGraph, Claude, MCP tools, RAG, structured extraction, call analytics and scoring. Evaluation sets for every prompt or model change, guardrails, deterministic outputs, logging and monitoring. 5. Voice and call AI Real-time voice agents (LiveKit, Pipecat, Twilio, SIP/PBX integration), call QA and analytics from recordings, CRM hand-off. 6. GPU infrastructure Linux, Docker, queues (NSQ), Prometheus/Grafana with alerts, FLAC archiving, human review tools for training data. HOW I WORK • Evaluation first: before changing a model or a prompt, we agree on a test set and a metric. You see the numbers before and after. • Short written updates, async-friendly, CET time zone, 0–4 h response time. • Fixed scope or hourly, your choice. Small paid audit first if you prefer to start low-risk. BACKGROUND 20 years in software. Co-founder and CTO of a survey research platform (web plus an Android field app, 1M+ questionnaires, 80+ languages), wrote ~97% of its backend. Technical lead for 6 years on a platform with its own crypto nodes and automated payouts. Earlier: head of IT for a multi-branch company, grew the team from 3 to 15. STACK AI/ML: PyTorch, Hugging Face, NVIDIA NeMo, CTranslate2, vLLM, LangGraph, Claude API, MCP Backend: Python, Go, PHP, MySQL, PostgreSQL, REST, NSQ Infra: Linux (Debian), Docker, Prometheus/Grafana, GPU servers (RTX PRO 6000 Blackwell, A6000, RTX 3090) Send a short audio sample, a model name or a one-paragraph task description, and I will tell you what result is realistic and what it will cost.