Skip to academic profile
SYLVESTER KACZMAREK

Sylvester Kaczmarek

Academia

Advancing trustworthy intelligence through research, teaching and experimentation.

I work across AI, cybersecurity and space systems, asking how intelligent systems can remain controllable and how their behaviour can be verified.

Academic appointments and teaching

  • OPITProfessor
  • Kepler Space UniversityAssociate Professor
  • University of OxfordTutor
Sylvester Kaczmarek
Research · Teaching · Academic collaboration

Research vision

Questions worth testing

My research agenda for the next five years centres on the relationship between capability, control and evidence. I want to understand what makes an intelligent system dependable, and how another researcher can test that claim.

AI also creates new opportunities for scientific discovery: identifying connections across disciplines, generating hypotheses that researchers might otherwise overlook, and helping design experiments that resolve uncertainty. I want to investigate which of these methods genuinely produce new knowledge and how their findings can be independently evaluated.

As AI takes on more routine research tasks, I place particular value on scientific judgement, original questions and experimental verification. Released code, datasets, evaluation protocols and reproducible experiments are central research contributions in this agenda.

Controllable and accountable AI

Develop methods for verifying that AI agents remain within delegated authority and that interventions reliably constrain their effects.

  • When is a cancellation demonstrably complete?
  • How should independent control operate across concurrent agents?

Intended outputs Experimental protocols, intervention benchmarks and auditable evidence linking permissions, actions and observed outcomes.

Trustworthy intelligence for physical systems

Investigate how autonomous and neuromorphic systems can maintain useful performance within explicit safety, energy and operational limits.

  • Which faults invalidate an assurance claim?
  • What can a spacecraft verify with limited telemetry?

Intended outputs Reproducible fault campaigns, hardware measurements and evaluation methods connecting computational behaviour to physical consequences.

Verifiable scientific discovery

Develop approaches to AI-assisted research that make scientific claims traceable to inspectable methods, data and experiments.

  • Which evidence enables independent replication?
  • How can an audit detect unsupported conclusions?

Intended outputs Reusable datasets where appropriate, provenance records, replication studies and evaluation protocols with explicit uncertainty and limitations.

Open research questions

From a question to an experiment

Seven starting points for research conversations. Open a question to see a possible method, contribution and relevant skills.

These are research directions for discussion. Specific projects, supervision arrangements and funding would need to be agreed with the relevant institution.

AI controlHow can an agent's halt be independently verified?

A cancellation experiment must account for outstanding work in tools, queues and external services.

Research question
Which independent observations support a bounded claim that consequential actions ceased by a stated deadline, within a specified system boundary?
Possible method
Instrument a sandbox's action boundary separately from the agent. Cancel runs at varied execution points and introduce delayed acknowledgements. Observe effects for a predeclared period after cancellation, checking timestamped traces for retried and already committed actions.
Potential contribution
A repeatable halt-verification protocol, with measured delays, observation gaps and explicit limits on the claim.
Relevant skills
Distributed systems, Python or TypeScript, experimental design.
AI controlCan control survive simultaneous agent failures?

Concurrent-agent experiments can test whether shared monitoring and intervention remain effective as alerts accumulate.

Research question
Which coordination designs preserve timely intervention when several agents fail or request conflicting actions together?
Possible method
Compare centralised and distributed monitors in an isolated testbed. Vary agent count, alert timing and communication loss; measure missed interventions, recovery time and unnecessary shutdowns. Set intervention deadlines before testing.
Potential contribution
A benchmark and evidence about the operating limits of each control design.
Relevant skills
Concurrency, distributed systems, reliability engineering, statistics.
Scientific evidenceWhat evidence is sufficient for adaptive AI assurance?

An adaptive system provides a test case for studying when earlier evaluation evidence remains relevant.

Research question
Which observable changes should trigger renewed evaluation, restricted operation or withdrawal of an assurance claim?
Possible method
Define claims and failure criteria before testing. Apply controlled model updates and distribution shifts; compare monitoring signals with independently measured failures and confidence intervals. Report false alarms and missed failures separately.
Potential contribution
An evidence framework connecting specific changes to proportionate revalidation and clearly bounded claims.
Relevant skills
Machine learning evaluation, assurance, statistics, systems engineering.
Physical systemsHow robust is neuromorphic intelligence under realistic faults?

Neuromorphic evaluation can connect inference quality with timing, energy consumption and fault response.

Research question
Which combinations of input disturbance and hardware fault most strongly affect useful performance and safe recovery?
Possible method
Run reproducible workloads on available hardware with controlled input perturbations and permitted fault injection. Compare against software baselines while measuring accuracy, latency, energy and recovery behaviour.
Potential contribution
A documented evaluation protocol and measurements that distinguish hardware effects from model behaviour.
Relevant skills
Embedded systems, spiking neural networks, instrumentation, statistics.
AI controlWhen do limited vulnerabilities create disproportionate system risk?

A controlled experiment can separate individual weaknesses from interactions between them.

Research question
Under what conditions does their joint effect exceed an explicitly defined additive risk baseline?
Possible method
Use a synthetic system with harmless failure scenarios. Run factorial experiments across individual and combined weaknesses, controlling exposure and measuring interaction effects with uncertainty estimates. State the risk measure and independence assumptions in advance.
Potential contribution
A reproducible method for detecting interaction effects and prioritising system-level mitigations.
Relevant skills
Security modelling, causal reasoning, experimental design, statistics.
Physical systemsWhat evidence supports autonomy beyond immediate human oversight?

Spacecraft scenarios allow investigation of autonomy under restricted communication, observation and computational resources.

Research question
What onboard evidence is sufficient to justify continued operation or a transition to a defined safe state?
Possible method
Combine simulation with hardware-in-the-loop experiments where available. Introduce telemetry gaps and sensor faults; compare onboard decisions with independently recorded state and delayed ground assessment.
Potential contribution
Evidence requirements and test procedures linking operational constraints to bounded autonomy decisions.
Relevant skills
Space systems, control, embedded software, uncertainty analysis.
Scientific evidenceCan AI-assisted scientific results withstand independent replication?

Replication studies can test how reliably another team reconstructs a result from its research record.

Research question
Which records expose unsupported inference, data leakage or sensitivity to undocumented analytical choices?
Possible method
Have independent reviewers reproduce the analysis from versioned data, code and protocols, then repeat a selected experiment on fresh data. Compare discrepancies and error detection across documentation and provenance conditions.
Potential contribution
An evaluated research-record specification and a replication protocol reporting discrepancies and unresolved uncertainty.
Relevant skills
Research methods, data engineering, statistics, domain expertise.

Teaching philosophy

Judgement through practice

I believe students should learn to formulate good questions, challenge assumptions and produce evidence that their conclusions are sound. My teaching connects theoretical foundations with practical experiments, reproducible software and real engineering problems.

AI-assisted tools can support this work. Students still need to explain their decisions, disclose how tools were used, verify outputs and recognise the limits of an analysis. I value an inspectable method, a well-designed test and an honest account of an unexpected result.

For example, in a software security assignment, students might use AI to propose a solution to an access-control vulnerability. They would then need to reproduce the problem, test the proposed solution against normal and adversarial cases, identify any remaining weaknesses and defend their conclusions using evidence. The assessment would emphasise their reasoning, experimental method and ability to verify the result.

  1. Frame the problem

    Define the question, assumptions and what a convincing answer would require.

  2. Build and test

    Connect theory to implementation, measurement and controlled experiments.

  3. Defend the evidence

    Make the reasoning reproducible and explain uncertainty and limitations.

Teaching portfolio

Current teaching and future curricula

Current teaching

OPIT

At OPIT, my teaching covers undergraduate and postgraduate study in the following areas.

  • Secure Software Development
  • Cryptography and Secure Communications
  • Parallel and Distributed Computing
  • Secure Large Language Models
  • Applied Machine Learning

Kepler Space University

I hold an Associate Professor appointment at Kepler Space University.

University of Oxford

Tutor · Harris Manchester College

I teach the course Worthy of Trust? AI and Autonomous Systems at Harris Manchester College as part of the University of Oxford's Inspiring Oxford programme.

Proposed courses

Subjects I would welcome developing with universities, adapted to their programmes, students and research priorities.

AI Safety Engineering
Hazard analysis, runtime intervention and experimental assurance.
Trustworthy Autonomous Systems
Bounded decision-making, fault tolerance and operational evidence.
Secure Agentic AI
Delegated authority, tool boundaries and adversarial evaluation.
Neuromorphic Computing
Event-driven models, embedded implementation and hardware evaluation.
Space Systems Autonomy
Onboard intelligence under communication, energy and safety constraints.
Reproducible AI Research
Experimental design, provenance, replication and uncertainty.

Research leadership and contributions

Work with an inspectable record

Selected work connects research questions with published analysis, experimental methods and engineering programmes. The Research page holds the wider publication record.

CETaS · The Alan Turing Institute

Future AI paradigms

As Principal Investigator on CETaS research into future AI paradigms, I co-authored the 2026 report The Next Frontier. The work examines emerging AI architectures, learning methods and computing hardware, with particular attention to security implications and UK research priorities.

Read the CETaS report

KRI

Independent verification

Through KRI, I develop research around trustworthy autonomy and independent evaluation. My agenda includes experimental audit trails, reproducibility and the assurance of AI-supported research workflows, connecting stated safeguards with observed behaviour.

Two public software implementations illustrate this approach. Guardian Agent Runtime tests whether delegated authority and safety constraints remain enforceable when AI agents behave incorrectly or adversarially. WorldState Check independently verifies whether autonomous actions achieved their intended outcomes using observable evidence and reproducible tests.

KRI research and open outputs · Guardian Agent Runtime · WorldState Check

OrbiSky · UK Space Agency-supported research

SkyYield

As CEO of OrbiSky, I lead work on SkyYield, which received UK Space Agency support for a feasibility study and payload concept for processing ZBLAN fluoride glass in microgravity. The engineering agenda centres on process control, experimental measurements and verification of material quality.

UK Space Agency programme announcement

Neuromorphic computing

Adaptation with assurance

My work examines how neuromorphic systems can adapt while retaining security and assurance requirements. It connects system design with the question of what evidence is needed to evaluate behaviour under operational constraints.

IEEE Computer paper

FDL LunarLab

Multimodal lunar research

As Research Lead for FDL LunarLab and a member of its steering committee, I help shape scientific evaluation of multimodal lunar intelligence, including provenance, uncertainty and reproducible experiments. I also co-authored research combining lunar observations for surface mapping and resource assessment.

MAPSIT research abstract (PDF)

Academic service

Contributing to the research community

  • NeurIPS-associated workshops. Completed four reviews for Machine Learning and the Physical Sciences (ML4PS) in 2025; accepted a further reviewer assignment for its 2026 edition.
  • ICLR workshops. Completed peer reviews for Foundation Models for Science and AI & PDE at ICLR 2026.
  • AAAI workshop. Completed peer review for the TrustAgent workshop at AAAI 2026.
  • EurIPS workshop. Completed peer review for Epistemic Intelligence in Machine Learning (EIML) at EurIPS 2025.
  • Journal peer review. Completed reviews for IEEE Transactions on Automatic Control, IEEE Transactions on Control Systems Technology, Wiley's Engineering Reports and the Journal of the British Interplanetary Society, including revised-manuscript review for IEEE TAC.
  • Conference service. Accepted an appointment to the Technical Programme Committee of IEEE SysCon 2026.
  • Research programmes. Faculty contribution to Frontier Development Lab in 2018, documented in the programme history.

Scientific evidence and advisory

Evidence for public scrutiny

Selected published submissions to UK parliamentary committees. Each addresses how research, technical evaluation and independent evidence can inform decisions about emerging technologies.

  • AI and copyright

    Technical evidence on enforceable rights reservation, data provenance and auditable transparency for AI developers.

    House of Lords · Communications and Digital Committee

  • Research infrastructure

    Evidence on research testbeds, digital infrastructure, reproducibility and long-term access to scientific capabilities.

    House of Commons · Public Accounts Committee

  • Low-energy computing

    Evidence on neuromorphic computing, realistic system-level energy measurements and independent benchmarking.

    House of Commons · Science, Innovation and Technology Committee

  • UK engagement with space

    Evidence on security, resilience and verifiable autonomy for research and operations in the space sector.

    House of Lords · UK Engagement with Space Committee

Students and early-career researchers

Bring a question and a way to test it

I welcome enquiries from undergraduate, master's and doctoral students interested in the questions above. Useful starting points include independent replication studies, benchmark development, experimental software and carefully scoped engineering investigations.

A short research note is a good introduction. Describe the question, why it matters, the method or baseline you would test, and the skills and time you can bring. Include your programme or institution and any requirements for formal supervision.

Formal supervision, studentships and institutional projects depend on the relevant university's arrangements and available capacity.

Discuss a research idea

Academic collaboration

Research and teaching in partnership

I welcome discussions with universities and research institutions interested in interdisciplinary research programmes, new courses, experimental collaborations and academic appointments.

Back to the top