AI Research & Engineering
Generative video · Archival film restoration · Agentic systems in production

🎓 M.Sc. in Artificial Intelligence — Wrocław University of Science and Technology · GPA 5.0/5.0, Highest Distinction

First-author work on restoring century-old film — two papers under review, plus a deployment with two Polish cultural institutions.

A recurrent transformer that predicts a soft defect mask and propagates it through time — reasoning explicitly about where footage is damaged and how severely, instead of reconstructing frames blindly.

• State-of-the-art perceptual quality on real archival benchmarks, outperforming BasicVSR++, RTN, DeepRemaster and MambaOFR
• 6.6M parameters at 0.35 GB — the most memory-efficient model in the comparison
• Multi-scale Dilation Pyramid MaskNet with direct mask supervision and AdaLN-Zero degradation conditioning
• Trained multi-node on HPC clusters (SLURM, DDP)

A degradation pipeline that models the full analog-to-digital chain — signal-dependent grain, mean-reverting (Ornstein–Uhlenbeck) gate weave, parametric scratches — paired with a large real-world archival benchmark.

• First synthesis method to jointly model all 7 analog artifact families with temporal coherence and a severity curriculum
• 81,576-frame benchmark curated from 30 public-domain films (1896–1918, Library of Congress)
• Models trained on the pipeline generalize measurably better to real footage, and expose failure modes of prior methods

🔒 Both papers are under review. Code and full results will be released on acceptance.

Self-initiated restoration collaborations with Narodowe Archiwum Cyfrowe (Polish National Digital Archive) and the Museum of Modern Art in Warsaw — restoring footage from their collections, with positive institutional feedback.

Research. Computer vision and generative modeling for video — temporal consistency, detail reconstruction, and degradation modeling for archival footage.

Production ML. At Grid Dynamics I built a hierarchical multi-agent system (Google ADK) for automated B2B sales research, and a production-grade Langfuse observability suite with full-lifecycle tracing, prompt versioning, and a dual-evaluator framework. Cut inference cost 30% via prompt caching and instance selection, and latency 15% via async agent execution. Shipped a config-driven agent starter-kit that became the internal standard, taking new agents from days to hours to production.

End-to-end. A full-stack background (React, Angular, NestJS, Spring Boot) means I don't just train models — I ship them as scalable, production-ready systems on GCP and Azure.

• 🎤 ML in PL 2025 — presented Generative Image Inpainting with Self-Supervised Learning (student workshop)
• ☁️ Google Cloud Associate Cloud Engineer
• 🧠 Deep Learning Specialization — DeepLearning.AI
• 🇬🇧 IELTS C1 — Academic English

mikjas.com · mikjas02@gmail.com · LinkedIn