Sofia, Bulgaria
Ivan Zdravkov
AI researcher who ships what he studies.
Doctoral researcher in adaptive LLMs, Bulgarian Academy of Sciences, from October 2026. Pre-registered experiments, published with real numbers. 20+ years turning findings into systems that run in production.
Research
Doctoral research, and the discipline that already runs in public
Adaptive large language models for social robotics through continual learning and knowledge retention mechanisms.
Every experiment on this site is pre-registered, with a kill-gate threshold fixed before any code runs, and published with the same weight whether the gate passes or fires. Frozen plans and results · Research notes
What I do
Systems that keep working after the demo
-
Applied AI in production
Retrieval-augmented systems grounded in your own data, document processing pipelines, anomaly detection over structured data.
-
Backend systems and integrations
PHP/Laravel and Python services, API integrations, systems built to run for years.
-
Privacy-preserving architecture
Masking layers before any model access, locally operated open models, GDPR and EU AI Act alignment.
How I work
Four principles
- Every output can be traced back to the data and the code that produced it.
- Pipelines are reproducible: same input, same result, on your hardware.
- No vendor lock-in: open models behind an abstraction, swappable without rewriting the pipeline.
- Honest reporting, including what did not work and why.
Selected work
Projects I can talk about
-
Applied AI
AI automation inside a cross-border logistics operator
Designing and running the AI integrations across a logistics company's processes: support triage, financial reconciliation, and grounded access to internal data.
-
Risk modelling
Acumen Risk: turning field reports into risk signals
A year of work on a risk management platform whose users include teams at BP and OMV, building the models, the algorithms and the signals behind its risk assessment.
-
Applied research
GreenDC: optimisation algorithms for data centre energy
Implementing the optimisation algorithms for an EU-funded project on forecasting and reducing energy consumption in a telecom operator's data centre.
-
IoT
Cloud portal for fog-based active protection systems
A portal that collects telemetry from and sends commands to security fog systems connected to customers' alarms, built for an Italian manufacturer.
Selected projects delivered through AFC EOOD, and earlier work as an employee.
Writing
Recent posts
-
Catastrophic forgetting, measured
A small network lost 12.3 points of accuracy learning a second task, real but short of my registered threshold for severe. EWC recovered 98% of it.
-
Blackwell quantisation on my own card
Q4_K_M more than doubled throughput and cut VRAM by two thirds on my RTX 5070 Ti, staying 98% similar to full precision. The interesting part was where that 2% lived.
-
What the EU AI Act actually requires of a masking layer
The Act never says masking. Three articles, on data, logging and human oversight, each describe a property a masking layer either has or does not.
-
Your model is not the product. The pipeline around it is.
Swap the model and almost nothing changes. Swap the masking layer, the permissions, or the monitoring, and the system stops running or starts leaking.
Contact
Start with a conversation
Every engagement starts with a conversation. Describe the problem you are trying to solve and I will come back with scope, approach and a quote.
- Email [email protected]
- LinkedIn linkedin.com/in/izdravkov
AFC EOOD, Bulgaria, founded 2004