Mykhailo Z.

Project overview:

Automated real-time transcription of interviews to generate official protocols in a secure environment. On-premise (air-gapped) deployment ensuring maximum security and data privacy. Core Model: Python, OpenAI Whisper, Pyannote, Docker, on-premise deployment Orchestration: Custom system for real-time processing (voice detection + chunking + transcription). Supports up to 10 simultaneous sessions. Fine-tuning Pipeline: Created a pipeline for periodic model updates using client-provided datasets (edited transcripts). Focused on adapting to (local dialect) and low-quality audio. Metrics: Used WER (Word Error Rate) and CER (Character Error Rate) to validate model performance. Deployment: On-premise (Air-gapped). All components are deployed locally to ensure maximum security and data privacy.

Responsibilities:

• Designed and built custom real-time processing system: voice detection (Pyannote) + chunking + transcription pipeline.
• Implemented batch optimization, buffer tuning, and custom VAD logic for real-time Whisper-based recognition.
• Created fine-tuning pipeline for periodic model updates using client-provided datasets (edited transcripts).
• Managed model evolution: OpenAI Whisper Medium → Large → Turbo.
• Built automated QA pipeline using WER/CER metrics with auto-retraining triggers.
• Deployed all components locally on air-gapped infrastructure (via RAY.IO, Docker, vLLM).

Achievements:

• A production-ready system (active for 3+ years) that generates real-time protocols from microphone input, resilient to background noise and street recordings. • Supports up to 10 simultaneous transcription sessions (with different numbers of users per session). • Reduced Whisper latency from ~4s (Medium model) to 1.05s (Turbo) while maintaining high accuracy in noisy environments. • Production quality thresholds: WER < 7%, CER < 7% with automated re-training when exceeded. • Successfully adapted recognition for local dialect and low-quality audio sources.

Technology stack:

Project overview:

Voice agent integration for a financial services company with a focus on mobile stability. Focus: Integrated AI agents with telephony infrastructure. Solved architectural challenges regarding vendor integrations. Performance: Focused on maintaining high communication quality over mobile networks.

Responsibilities:

• Designed and implemented real-time voice communication architecture: Asterisk PBX ↔AudioSocket ↔ OpenAI Realtime API.
• Optimized latency by switching from 16-bit PCM to 8-bit codecs (G.711/μ-law).
• Configured intermediate server routing for optimal network paths.
• Implemented RTP over UDP for production telephony to minimize delays.
• Integrated AI agents with telephony infrastructure, resolving vendor-specific challenges.
• Built and maintained voice agent flow: call handling, speech recognition, LLM processing, TTS response.

Achievements:

• Reduced voice latency from 200–250ms to 60–80ms through codec optimization and routing. • Successfully deployed real-time voice AI agent over unstable mobile networks in Uganda. • Built a production-grade bridge between Asterisk PBX and OpenAI Realtime API. • Achieved stable call quality using UDP protocol with minimal buffering layers.

Technology stack:

Project overview:

Knowledge base system for medical device documentation with semantic search capabilities. Pipeline: Web scraping of manufacturer manuals for specified medical devices → chunking → indexing with metadata → storage in vector database. Core Functionality: On query, retrieves relevant documentation and specifications for a given medical device.

Responsibilities:

• Designed and implemented web scraping pipelines to collect manufacturer manuals and device documentation.
• Developed chunking and indexing strategies with metadata tagging for accurate retrieval.
• Configured ChromaDB as vector store with metadata filtering for device-specific queries.
• Integrated all-MiniLM-L6-v2 embedding model for semantic search capabilities.
• Built RAG pipeline using LangChain with Llama 2 as the generation model.
• Set up distributed processing with Ray.io for scalable document ingestion.

Achievements:

• Built a comprehensive knowledge base covering manufacturer manuals and specifications for medical devices. • Achieved high retrieval accuracy through optimized chunking strategies and metadata enrichment. • Reduced query response time to sub-second levels through embedding model selection and vector store optimization (as mixed search via cosine similarity and metadata usage).

Technology stack:

Project overview:

Automated document processing system for extracting structured data from diverse file formats and translating into target languages. Input: PDF, images, DOCX, TXT and other file formats. Core Pipeline: File ingestion → OCR extraction (Qwen2.5-VL) → structured JSON output for rendering → translation to target language (Gemma 3).

Responsibilities:

• Designed end-to-end document processing architecture: ingestion → OCR → structuring → translation.
• Implemented OCR extraction using Qwen2.5-VL (Qwen3-VL later) served via vLLM for high-throughput inference.
• Built translation module using Gemma 3 (GemmaTranslate later) served via Ollama for multi-language support.
• Developed structured JSON output schema for consistent rendering across document types.
• Configured Kafka message queue for asynchronous document processing and load balancing.
• Set up distributed orchestration with Ray.io for parallel processing of large document batches.
• Containerized all services with Docker for reproducible deployment.

Achievements:

Scalable pipeline that processes heterogeneous documents into structured, translatable output with support for multiple target languages.

Technology stack: