GitHub LinkedIn HuggingFace HuggingFace

Dennis Khylkouski

Cybersecurity @ Cerebral Valley. I work on security research, agent infrastructure, and model optimization.

AI Systems

I like making models and agents run faster, cheaper, and with fewer artificial constraints. I got an uncensored Qwen 3.6 35B running on a single H100 around 3x faster than baseline, moving from roughly 100 tok/s to 250+ tok/s with prefix caching, faster attention and MoE backends, and DFlash speculative decoding with a matched draft model. I also fine-tuned Qwen on cybersecurity datasets for vulnerability discovery and exploit analysis.

I built an agent orchestration system for running thousands of coding agents simultaneously, and built Baremodel, an uncensored model API for coding agents and automation workflows. I have also experimented with watermark robustness, including removing SynthID image watermarks through diffusion reconstruction and denoising.

Security Research

I have found and responsibly reported vulnerabilities across 10+ startups and public-sector systems. The work has included issues in UC Berkeley systems, AI and Kubernetes infrastructure, Carnival Cruise systems affecting passenger data, and exposed admin dashboards across maritime infrastructure tied to $40B+ in shipping assets.

Some of my research comes from reverse engineering real systems. I reverse engineered Bird's scooter API and mapped about 1,000 live scooters in San Francisco.

Competitions

Recruiting Intelligence

One thing I have been working on is ranking people by evidence instead of what they say about themselves. The idea is that competition results, commit history, and writing are more reliable signals than a resume, and the question is how to collect and weight them.

That turned into Metric, which I co-founded with my brother. It pulls signals from competitions, repos, talks, and private or paywalled sources, then outputs ranked candidates with scores and a breakdown of why they placed where they did. We backtested against real outcomes and got 93% accuracy.

Projects

Miscellaneous