Builder mindset
Own the path from an unclear problem to a shipped, operable system.
AI SECURITY · COMPLIANCE · APPLIED RESEARCH
Sebastian Haas
Always a nerd. Always building.
I build and test AI systems, from GPU kernels to autonomous agents. My work connects adversarial evaluation, compliance engineering, and applied research in the defense industry.
Certifications
- Certified Red Team Operator II (CRTO II)
- Certified Red Team Operator (CRTO I)
- Practical Network Penetration Tester (PNPT)
- Practical AI Pentest Associate (PAPA)
- Security Analyst Level 1 (SAL1)
- Blue Team Level 1 (BTL1)
- BSI IT-Grundschutz-Praktiker
- BSI IT-Grundschutz-Ersthelfer
- CompTIA Cloud+
- CompTIA Security+
AI security
Attack systems early, find edge cases, and engineer controls that survive.
Compliance
Translate obligations into technical controls, traceability, and evidence.
Research
Test assumptions through reproducible experiments and technical writing.
Selected work
All projects →Custom GPU kernels, speculative decoding, and a Rust serving layer. Months of building, profiling, and following failures through the full inference stack.
Explore the engineering →
35B-A3B code generation149.8 tok/sSelected DFlash peak on two 16-GiB Radeon Pro VII GPUs. Q4_0 / Router257.
Qwen3-8B · 32k decode4.1×14.9 to 60.5 tok/s with q8 KV. Compared with upstream b10524's conversion path.
Qwen3-8B · concurrent serving750 tok/sAggregate decode across two MI50 instances, eight streams each. Twice the single-card throughput.
Lightning 30B-A3B · 32k decode+84%49.4 to 91.0 tok/s versus upstream. Q4_0, q8 KV, with six expert blocks on CPU.
Separate development benchmarks. Performance depends on model, context, and workload. Methods, baselines, and limitations →
Let’s talk AI, security, and what’s next.
I’m interested in engineering collaborations and research on agent security, adversarial evaluation, and practical AI controls. Have a specific problem or experiment in mind? I’d like to hear about it.
hello@sebastianhaas.dev →