Founder & CEO, Sober Agentic Infrastructure, Inc. · Operational Sobriety for the Agentic Era. Founder & Creator, VAREK · AI Pipeline Programming Language.

I'm Ken — physician turned founder, based in Dallas-Fort Worth.

Founder & CEO of Sober Agentic Infrastructure, Inc. (soberagents.ai), building verification infrastructure for agentic AI. Creator of VAREK.

Medicine doesn't deploy systems that are "usually right" — it requires verification before action. I'm applying that standard to AI agents.

Currently building: deterministic runtime verification · compile-time pipeline safety · pre-execution policy decisions

Background: clinical medicine → AI safety infrastructure

varek-lang.org · github.com/kwdoug63/varek · MIT License

VAREK is an open source programming language designed from the ground up for AI/ML pipeline engineering. It replaces the Python + YAML + JSON Schema + Bash fragmentation that plagues production AI systems with a single typed, compiled language where pipelines are first-class citizens — and verifies what the agents built on those pipelines are allowed to do.

• Statically typed — Hindley-Milner inference extended with tensor shape tracking
• LLVM-compiled — 10–40x faster than CPython for compute-heavy pipeline workloads
• Pipeline declarations — type-checked at compile time, not discovered at runtime
• Full interop — import python::numpy, import python::torch — no rewrites required
• v1.0 language baseline — 659 tests passing · 7-module standard library · 261 functions · package manager
• v1.5 Warden runtime — kernel-level interception via seccomp-unotify · sub-microsecond policy decisions · three-state ALLOW / DENY / UNKNOWN return with symmetric suppression
• v1.6 plan verification — the entire planned action graph is verified before execution begins · an unsafe plan is refused before the agent's first action · composes with the per-action runtime as a second deterministic layer

Deterministic AI safety in depth: verified at compile time in the language, across the whole plan before execution, and per action at the kernel.

soberagents.ai

We build model-agnostic, neuro-symbolic guardrails that sit between AI agents and critical infrastructure.

• Deterministic Safety: An SMT decision procedure mathematically blocks unsafe actions before they execute.
• Plan Verification: A simulation layer that validates agent plans before execution.
• Liability Mitigation: Closing the "Trust Gap" that prevents Fortune 500s from deploying autonomous agents at scale.

My approach to AI safety is not theoretical; it is forged in environments where failure is not an option:

• Medical Precision: Doctor of Medicine (MD), USC Keck School of Medicine. I understand systemic failure and the necessity of "clinical" guardrails in autonomous systems.

• Language Design: Built VAREK from first principles — lexer, parser, type system, LLVM backend, standard library, package manager. 659 tests. 13,006 lines. One year.

• Hardware Innovation: Lead Author of US Patent 12,037,144 B1 (Space Railway). I build systems that rely on the laws of physics and logic, not probability.

• Software Innovation: Lead Author of a three-patent portfolio forming a vertical kernel-to-formal-methods stack for agentic AI safety:

  • US Provisional 64/006,104 · Filed Mar 15, 2026 — Deterministic Safety Guardrails and Outcome Verification for Agentic Systems
  • US Provisional 64/059,592 · Filed May 7, 2026 — Kernel-Level Interception Architecture for Autonomous Agent Containment
  • US Provisional 64/062,549 · Filed May 11, 2026 — Compositional Pre-Execution Verification of Agent Action Plans

  I build systems that rely on the laws of math and logic, not probabilistic guesswork.

• VAREK: Python (reference implementation) · LLVM IR via libLLVM-20 · Hindley-Milner type system
• Sober Agentic Core: Python 3.11+ · Rust (Logic Kernel) · FastAPI · Docker
• Formal Methods: SMT decision procedures · First-Order Logic (FOL) predicates · Symbolic Verification
• Predictive Math: NumPy/SciPy for latent-space bounding and action-validation

• VAREK: varek-lang.org
• Sober Agentic: soberagents.ai
• LinkedIn: linkedin.com/in/kwdoug63
• Email: kenneth.douglas@soberagents.ai

"We don't need smarter models; we need soberer infrastructure."