About Roman
Architect-level AI Engineer | Knowledge Engineering | 12+ years of experience
I solve the data chaos problem for companies that need reliable AI Systems. I transform unstable prototypes into high-performance architectures, specializing in Knowledge Distillation, Agentic workflows, and cost-optimized production deployments.
What I deliver:
- Production AI: From 2-week PoCs to scalable MLOps deployment.
- Cost Efficiency: Specialized in reducing LLM inference costs (up to 50% reduction in recent projects).
- Deep Domain Expertise: High-stakes industries including Pharma, Radiology (VLM), and Veterinary MedTech.
Core Technical Focus:
- AI Agents: Complex orchestration using LangGraph and autonomous workflows.
- Advanced RAG: Enterprise knowledge systems and intelligent corporate memory.
- Voice & Audio: Clinical ASR and real-time speech-to-action solutions.
- Computer Vision: Medical/Veterinary imaging analysis using VLMs.
Key Achievements:
- Pharma: Designed an LLM retrieval platform that optimized costs by 45%.
- Healthcare: Built an autonomous Multi-Agent ecosystem for clinical automation (LangGraph + ASR + CV).
Based in Paris. Available for remote or hybrid missions.
English
Fluent
French
Conversational
Russian
Native or bilingual
Remote only
Primarily works remotely
Experience
- TN SystemsSenior AI EngineerDIGITAL AND ITSeptember 2025 - Today (9 months)Paris, FranceAutonomous Multi-Agent AI System for Veterinary ClinicsBuilt a full-stack AI ecosystem integrating RAG, Multi-Agent orchestration, Streaming ASR, and Multimodal Computer Vision to fully automate clinical diagnosis and documentation workflows.Key Contributions:•Multi-Agent Architecture: Designed orchestration logic (LangGraph) coordinating specialised agents (ASR, RAG, Radiology, Reasoning) into a unified diagnostic system.•Streaming ASR: Engineered a high-performance real-time ASR pipeline (fine-tuned model) to transcribe vet-client dialogues automatically.•CV/Multimodal (Radiology): Implemented X-ray analysis module based on self-supervised JEPA architecture — generates structured image descriptions and differential diagnoses.•MLOps & Performance: Optimised inference services including a Rust-based pipeline for minimal latency in real-time critical functions.•AI-Native Development: Leveraged Cursor and Claude Code to maximise engineering velocity across the full stack.Results:•↓ 50% appointment documentation time → doubled clinic throughput•↓ Diagnostic and documentation errors via structured reasoning and validated medical templates
- Vivanti SALead Applied AI EngineerApril 2024 - September 2025 (1 year and 5 months)BelgiumDigital Marketing & AI for Life Science / PharmaProject 1: LLM-Powered Knowledge Retrieval Platform (Pharma)• Architected and deployed a RAG-based knowledge retrieval system fine-tuned for pharmaceutical data, integrated with internal corporate platforms.• Designed full lifecycle pipeline: data ingestion, embedding storage, domain-specific fine-tuning, and cloud inference optimisation.• Implemented AI-driven report generation module powered by autonomous agents.Results: ↓ 70% manual preprocessing · ↓ 40% inference costs · Automated end-to-end generation of internal training materials.Project 2: AI Training Platform for Medical Representatives (LLM Simulation)• Engineered real-time interaction loop combining LLM-driven objection handling with lip-sync avatars.• Built LLM-powered agents generating dynamic, context-aware medical dialogues tailored to specific pharmaceutical products.• Delivered customisable scenario templates (therapeutic areas, scripts, complexity levels).Results: ↓ 80% manual training preparation · ↓ 50% cloud inference costs via quantization and caching · Automated scenario creation and reporting.Tech Stack: Python, LangChain, LangGraph, LlamaIndex, HuggingFace, Ollama, AWS, Kubernetes, Docker, FastAPI, CI/CD
- FinTech (High-Load Banking Platform)Lead ML EngineerNovember 2021 - March 2024 (2 years and 4 months)Strategic Tech Partner for Top-Tier Banking (High-Load Fintech)Project 1: Next-Best-Action (NBA) Platform & Feature Store• Designed core ML infrastructure serving 10M+ customers with real-time personalised recommendations.• Built production-grade Feature Store with strict point-in-time correctness for training and inference.• Implemented automated drift detection (data & concept drift) and retraining pipelines.Results: ↑ Sales conversion via personalised real-time financial product offers · Stabilised model performance in production.Project 2: HR Analytics & Workforce Intelligence Platform• Developed end-to-end AI solution for recruitment and HR lifecycle optimisation.• Built scalable pipelines for data ingestion, feature engineering, and model deployment integrated with internal HR systems.Results: ↓ Time-to-hire via automated candidate screening · Churn prediction model identifying at-risk employees with 75% precision.
Recommendations
Be the first to recommend Roman
Help this freelancer shine by sharing your experience working together.
These freelancer profiles also match your criteria
Agatha Frydrych
Backend Java Software Engineer
4.7
(3)
2
Baptiste Duhen
Fullstack developer
4.6
(4)
5
Amed Hamou
Senior Lead Developer
4
(2)
7
Audrey Champion
Web developer
4.3
(3)
4
Education
- M.Sc. in Applied Mathematics and Computer SciencePerm State National Research University1997M.Sc.
- Machine Learning and Data AnalysisMIPT Yandex2019Machine Learning and Data Analysis
Certifications
- AI for Medicine Specializationdeeplearning.ai2025