Skip to content
View jersonboydmilan's full-sized avatar
🎯
Focusing
🎯
Focusing

Achievements

Achievement: Pair ExtraordinaireAchievement: Pull SharkAchievement: Quickdraw

Achievements

Achievement: Pair ExtraordinaireAchievement: Pull SharkAchievement: Quickdraw

Block or report jersonboydmilan

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
jersonboydmilan/README.md

Jerson Boyd Milan

AI systems, autonomous agents, and Physical AI — local-first, practical intelligence.

Website CUDA-Q contributor Merged upstream

I build agent systems that can act with real capability without handing the agent, or an attacker who compromises it, unrestricted power. Most of my time goes into authority and safety architecture for autonomous agents, and into upstream fixes for the open-source tools I build on.

Merged upstream

  • NVIDIA CUDA-Q · #5459: documented exp_pauli, givens_rotation and fermionic_swap, and fixed a sign error in the applyExpPauli doc comment that every simulator backend inherits. Verified numerically against the gate sequences CUDA-Q executes.
  • NVIDIA CUDA-Q · #5609: documented how custom operations take qvector targets in Python vs C++. Its review led CUDA-Q's architect to state the language's operation-overload rules (later applied in #5652).
  • NVIDIA CDF Simulation · #8: Surface Evolver tests now skip, instead of erroring, when the solver is registered but its binary is missing. Verified with the repo's own CI runner and skip policy.
  • AgentKit · #71: x-/X- extension fields in the Agentkitfile, so values can be defined once and reused through YAML anchors.

Filed at the CUDA-Q architect's request: Python bridge bugs #5646 and #5647, the latter now fixed on main.

Featured projects

Sentinel
An authority-attenuating immune architecture for AI agents. Detects and contains compromised agents without giving defenders unrestricted authority.
Python · agent security

Capgate
A runtime authority boundary for autonomous agents: the agent proposes, the harness authorizes, the executor acts. Deny-by-default policy, non-transitive delegation, tamper-evident audit.
Python · MCP gateway · isolation

Agent-Skills-Compendium
An open registry of fully specified AI agent skills, with taxonomy, canonical schema, composer, governance model, and a fleet runner.
TypeScript · multi-agent systems

QUANTUM-CUDA
The analysis, runnable verification, and patches behind my CUDA-Q contributions. Every claim is checked against what CUDA-Q actually executes, not transcribed from comments.
Python · quantum computing

In review

Project Contribution Status
NVIDIA CUDA-Q Fixed the custom KrausChannel example under cudaq.apply_noise, which crashes as printed #5608
Made the Observe examples on "Executing Kernels" runnable on their own #5655
"Building Kernels" shows the state each example prepares instead of a bare call #5658
Single-target broadcast for custom operations, plain and adjoint, under rules confirmed by CUDA-Q's architect; verified on the real runtime #5649
AgentKit Every Agentkitfile validation error names its file and line, with a BuildKit source snippet #75
Microsoft Agent Governance Toolkit A 240-scenario red-team suite that scores real detector catch/miss rates on prompt-injection probes #4346
guardrail docker_container_guard: blocks docker run/create flags that break container isolation #52
RepoLens A --schema flag exporting the JSON Schema of the machine-readable report #5

How I work

  • Verify, don't transcribe. Every claim in a PR is checked against what the code actually does, using the real runtime, a reproducer, or a numerical check, and the script goes into the repo.
  • Small, reviewable changes. One concern per PR, history traced through the linked issues, and open questions left to the maintainers instead of settled unilaterally.
  • Honest scope. What I couldn't verify is stated in the PR, not glossed over.

Now

  • Designing authority boundaries for agents: deny-by-default policy, non-transitive delegation, tamper-evident audit.
  • Contributing to NVIDIA CUDA-Q: docs and Python bridge fixes, each verified against the real runtime.
  • Exploring Physical AI data pipelines and agent tooling.

Agent authority & permission boundaries · Multi-agent governance · Skill composition · Physical AI · Local-first AI

Pinned Loading

  1. cuda-quantum cuda-quantum Public

    Forked from NVIDIA/cuda-quantum

    C++ and Python support for the CUDA Quantum programming model for heterogeneous quantum-classical workflows

    C++

  2. Nvidia-Physical-AI-Data-Factory Nvidia-Physical-AI-Data-Factory Public

    Shell

  3. QUANTUM-CUDA QUANTUM-CUDA Public

    Analysis, runtime verification, and patches behind my NVIDIA CUDA-Q contributions — docs fixes, Python bridge bugs, and scripts that check every claim against what CUDA-Q executes.

    Python

  4. Sentinel Sentinel Public

    SENTINEL: an authority-attenuating immune architecture for AI agents. Detect and contain compromised agents without giving the defenders unrestricted authority.

    Python 5 2

  5. Agent-Skills-Compendium Agent-Skills-Compendium Public

    An open registry of 76 fully specified AI agent skills — taxonomy, canonical schema, composer, governance model, and a fleet runner that deploys hundreds of agents against one objective and rehears…

    TypeScript 6 3

  6. Capgate Capgate Public

    Capgate — runtime authority boundary for autonomous AI agents: the agent proposes, the harness authorizes, the executor acts. Deny-by-default policy, non-transitive delegation, tamper-evident audit…

    Python 5