Open Positions

About Qualion Intelligence

Qualion Intelligence is a research and engineering company building reliable, auditable, and grounded AI systems for regulated sectors, contexts where probabilistic confabulation, opaque decision-making, and non-deterministic behavior are not acceptable trade-offs.

Our approach rests on a single architectural commitment: auditable AI without generative models. Our systems produce deterministic outputs, maintain cryptographically verifiable decision trails, and submit every internal mechanism to causal isolation through factorial ablation. The methodology has been peer-reviewed twice and independently replicated three times across different research groups.

About the work

When you see what modern AI systems produce, do you wonder how a regulator, an auditor, or a safety-critical operator can ever trust them? That question is the starting point for our work.

We build neuro-symbolic systems that combine structured reasoning, embedding-based perception, and runtime hard-constraint enforcement to produce traceable, deterministic decisions. There is no large language model in the inference path. Every output can be reconstructed, audited, and challenged.

In upcoming phases, we are exploring tighter integration of neural networks and world models into the architecture. These are not replacements for the symbolic core; they are candidate components inside it, each chosen to preserve determinism, replay, and audit. Four directions are active or planned:

• Perception and encoding. Extending the existing frozen-encoder pattern beyond text into vision (frozen ViT, DINOv2, SAM2 for segmentation) and audio (Whisper encoder used as encoder only). Encoders remain frozen, deterministic given input, hashable, replayable.
• Learned forward dynamics for planning. Augmenting symbolic forward simulation with neural dynamics models the planner can query as a fast simulator. JEPA-style joint-embedding predictive architectures are of particular interest because they predict in representation space rather than generating observations, which fits the non-generative commitment.
• Learned heuristics for symbolic search. Replacing hand-designed search heuristics with learned priority functions, following the AlphaZero pattern: the search itself remains symbolic and inspectable, only the ordering is learned.
• Self-supervised models over the audit trail. The system already generates hundreds of thousands of audit-trail events. Training internal predictive models on this stream gives us anomaly detection, drift monitoring, and subsystem-activation calibration without compromising the chain itself.

We are preparing the system for enterprise pilot deployment in a regulated sector during the second half of 2026. Both roles exist to make this transition possible.

About the roles

We are hiring for two engineering roles, with multiple openings on each. The ML Research Engineer role takes the research codebase from internal use to a system that survives external evaluation. The AI Safety & Compliance Engineer role translates our auditable properties into documentation that survives external audit under the EU AI Act, ISO/IEC 42001, and sector-specific standards.

Both are full-time. Both can be based in Ljubljana or anywhere in the EU with a regular monthly on-site presence.

ML Research Engineer

We are hiring multiple ML Research Engineers.

Ljubljana, Slovenia · EU-Remote eligible · Full-time

About the role

As the ML Research Engineer on the pilot team, you will work alongside the founder to prepare the research codebase for external evaluation. The work spans evaluation infrastructure, deployment engineering, reproducibility tooling, and adversarial testing.

You will not be researching new architectures. You will be making a real, working, well-documented research system perform reliably under contact with a serious enterprise customer. This is the engineering substrate that determines whether the pilot succeeds.

Responsibilities

• Productionise the Multi-Projection Gate inference path (frozen encoder plus three linear projection heads, currently 1.2ms per decision) for sustained throughput and reliable serving.
• Rebuild the factorial ablation pipeline (3×2×2 design, 11,700 trials, bootstrap CIs, TOST equivalence testing) as a reproducible command-line tool with full provenance tracking.
• Implement continuous evaluation infrastructure against HarmBench, AdvBench, SimpleSafetyTests-style benchmarks, plus pilot-specific evaluation sets developed jointly with the customer.
• Own the cryptographic audit trail infrastructure: hash-chained decision logs, deterministic seed control, model artefact provenance, replay verification.
• Maintain and extend the adversarial paraphrase protocol (500-item bank, 5 evasion strategies) and the graduated perturbation evaluation harness.
• Profile and optimise the system tick loop (currently 0.49ms baseline for non-gate subsystems) to maintain real-time guarantees under load.
• Document the codebase to a level where external auditors can verify claims independently.

You may be a good fit if you

• Have 4+ years of production Python and PyTorch experience writing code that survives without you: typed, tested, profiled, documented.
• Have shipped embedding-based ML systems in production using sentence-transformers, frozen encoders, or learned projection heads.
• Have implemented or worked closely with contrastive learning systems involving auxiliary loss terms such as decorrelation, orthogonality, or margin-based penalties.
• Are fluent in statistical experimental rigor: factorial ANOVA, bootstrap confidence intervals, equivalence testing, multiple comparison correction.
• Practice reproducibility discipline by default: seed control, deterministic inference, model artefact hashing, environment pinning.
• Can read research papers and implement from them with minimal supervision.
• Communicate clearly in written English and are comfortable working with auditors, partners, and external collaborators.

Strong candidates may also have

• Experience with cognitive architectures such as Soar, ACT-R, LIDA, or any neuro-symbolic system.
• Background in adversarial robustness: GCG-style token attacks, paraphrase attacks, or systematic red-teaming methodologies.
• Work in out-of-distribution detection, calibration, or conformal prediction.
• Experience with ARC-AGI benchmarks or other structured-reasoning evaluation suites.
• Open-source contributions to ML evaluation, reproducibility tooling, or research infrastructure.

AI Safety & Compliance Engineer

We are hiring multiple AI Safety & Compliance Engineers.

Ljubljana, Slovenia · EU-Remote eligible · Full-time

About the role

The technical substrate of our system (factorial ablation, deterministic inference, cryptographic audit trails) produces what we believe is one of the strongest auditability positions in current AI research. Your role is to translate that substrate into the format that real auditors, real risk officers, and real notified bodies expect to see.

Without this work, the technical results do not reach a customer. With it, the system becomes deployable in sectors where most AI systems structurally cannot operate.

Responsibilities

• Map our system's properties to EU AI Act Articles 9–15 (risk management, data governance, technical documentation, record-keeping, transparency, human oversight, accuracy and robustness) and produce the corresponding documentation.
• Prepare conformity assessment artefacts for high-risk classification under Annex III, including the technical file expected by notified bodies.
• Align our practices with ISO/IEC 42001 (AI management systems) and ISO/IEC 23894 (AI risk management).
• Conduct GDPR Article 35 Data Protection Impact Assessments for any data-touching pilot component.
• Own the customer-facing audit narrative: what we can prove, how we prove it, what residual risks remain, and what compensating controls exist.
• Interface with the customer's legal, risk, and compliance functions; defend our methodology under technical and regulatory review.
• Maintain our internal compliance posture, including post-market monitoring procedures for pilot deployment.

You may be a good fit if you

• Have 4+ years in AI governance, AI compliance, or regulated-industry AI deployment, and have shipped real documentation for real auditors.
• Have deep working knowledge of the EU AI Act: high-risk classification, Annex III, conformity assessment procedures, post-market monitoring, and the August 2026 obligations.
• Are technically literate: you can read ML papers and understand what a factorial ablation result actually demonstrates and what it does not.
• Practice documentation discipline: your writing survives hostile audit review, and you know the difference between a claim, evidence for a claim, and a falsification condition.
• Have working experience with one or more of: MDR / ISO 13485 (medical devices), ISO 26262 / ASPICE / TISAX (automotive), Basel / SR 11-7 (finance), ISO/IEC 27001 (information security), GDPR Article 35 DPIA practice.
• Communicate clearly in written and spoken English. German is a strong plus given pilot context.

Strong candidates may also have

• Direct experience interacting with notified bodies, regulators, or external auditors.
• Authored technical documentation for regulatory submissions: CE marking, FDA 510(k), or equivalent.
• Background in structured risk methodology: FMEA, STPA, bow-tie analysis, or equivalent hazard analysis.
• Familiarity with the NIST AI Risk Management Framework, OECD AI Principles, or sector-specific AI guidance.
• Legal training (LL.M., bar admission) combined with technical AI literacy. This profile is rare and highly valued.

Logistics

How we're different

We are a small, technically dense team building something that does not exist elsewhere: a cognitive system engineered from the ground up for auditability. We work as a single coherent unit on one focused research and engineering effort, not a portfolio of unrelated products. We treat AI research as empirical science, with the rigor that implies. We value clarity of thought and writing over surface impressiveness.

We will share our published technical work with shortlisted candidates ahead of any technical interview.

How to apply

Send a standard application by email. Both Slovenian and English are welcome, though English is preferred for the application itself.

Please include in your application

• Your CV or LinkedIn profile.
• A short cover letter (no specific length requirement) specifying which position you are applying for.
• Links to relevant work: repositories, publications, documentation, deployed systems, or anything else that best represents your experience.
• Your earliest possible start date and salary expectation.

We respond to every application within ten working days.

Send to jobs@qualion-intelligence.com
Suggested subject line: Application for [Role Name], [Your Name]