ML Systems & Agent Evaluation Engineer
San Marcos, Texas · montoyaraul34@gmail.com
GitHub Projects · Hugging Face · Limen Neural
Building reproducible model systems and coding-agent evaluation infrastructure on hardware I can own.
I build Rust/Python/CUDA tooling for model systems, coding-agent evaluation, and reproducible ML experiments: open-weight checkpoint analysis, MoE compression experiments, engineering-trajectory datasets, agent review/repair loops, and neuromorphic runtimes.
My current hiring focus is ML systems, coding-agent / eval infrastructure, model tooling, and research engineering. Neuromorphic computing is a specialist research track rather than a prerequisite for understanding the rest of the portfolio. Pursuing AI Engineering @ WGU.
| Area | Evidence |
|---|---|
| Agent systems & evals | Coding-agent trajectories, synthetic-data validation, multi-agent worktree experiments, deterministic review/repair gates |
| Model systems | Raw checkpoint parsing, MoE routing analysis, GGUF/Safetensors tooling, route-preserving quantization experiments |
| Systems engineering | Rust-first CLIs/runtimes, Python data/eval tooling, CUDA/GPU validation, CI, reproducible artifacts |
| External OSS | Merged xAI/Grok provider work in agent-afk, including API-key + OAuth support and a later compatibility fix |
| Public artifacts | Hugging Face datasets, telemetry corpora, Grok-1 research packs, manifests, experiment reports |
| Research specialty | SNN / neuromorphic runtimes, telemetry-driven control, FPGA-oriented export and validation |
Measurable results · reproducibility · provenance · fail-closed validation · explicit scope boundaries · humans merge
agent-afkPR #1019 — first-class xAI/Grok provider — merged. Added API-key and SuperGrok OAuth paths, provider selection, credential handling, docs, and tests in an external codebase.agent-afkPR #1242 — xAI OAuth CLI-version compatibility — merged. Fixed proxy-version negotiation with validated overrides/fallbacks and project-wide validation.
These matter to me because they test a different skill than owning my own repositories: understanding another project's architecture, satisfying its contracts, and getting work accepted upstream.
- Hugging Face portfolio — public model, dataset, telemetry, and research artifacts.
- Grok-1 GOZ1 research packs — packaged Grok-1 quantization research artifacts with manifests/provenance.
- Agentic Coding Trajectories — observable coding-agent trajectories for evaluation/training research.
- Spikenaut SNN Telemetry — machine telemetry and neuromorphic research data.
| Project | What it demonstrates |
|---|---|
| xai-dissect | Rust CLI for read-only structural analysis of open Grok-1 checkpoints: raw shard parsing, tensor inventory, MoE expert/routing maps, statistics, and machine-readable manifests |
| synthetic-factory | Synthetic-data generation and curation with schemas, provenance, strict audit gates, failure-oriented fixtures, and public Hugging Face release plumbing |
| operation-prometheus | Software-engineering trajectory forge: issue → implementation → review → fix → validation → merge datasets for coding-agent research |
| grok-ozempic | Real Grok-1 compression experiments focused on expert precision, routing fidelity, residual drift, and reproducible comparison reports |
| corinth-canal | End-to-end Rust reference pipeline spanning telemetry encoding, spiking hidden state, projection, GGUF/Safetensors model inspection, MoE routing, and SAAQ validation |
| Spikenaut-SNN | From-scratch SNN research runtime and export target for telemetry / neuromorphic experiments |
I also maintain modular libraries under Limen Neural, including projects such as neuromod, axon-encoder, nir-rs, synaptic-mesh, and brainstem-daemon.
Languages
Rust · Python · Julia · CUDA C/C++ · SystemVerilog
ML / model systems
PyTorch · Hugging Face · MoE · GGUF · Safetensors · LoRA/QLoRA · quantization experiments · SNN / neuromorphic systems
Engineering / infrastructure
GitHub Actions · Docker · Linux · self-hosted GPU CI · structured JSON/JSONL/Parquet artifacts · reproducible experiment manifests · automated review gates
Local research hardware
NVIDIA RTX 5080 16 GiB · AMD Ryzen 9 9950X · Fedora Linux
I use coding agents heavily, but I do not treat generated output as ground truth.
Typical flow:
issue / hypothesis
↓
isolated branch or worktree
↓
implementation
↓
local tests / lint / benchmark / experiment gate
↓
pull request
↓
independent review + repair
↓
human merge / release decision
Agent assistance is explicitly attributed in PRs, commit trailers, experiment notes, or release provenance where appropriate. I care about being able to explain and reproduce the result—not about pretending the tools were not used.
The repositories compose into five broader programs. Recruiters do not need to read all five to understand the flagship work above; these links are the deeper map.
- Autonomous Software Engineering — coding-agent worktrees, evaluation, engineering trajectories, review/repair loops.
- Frontier Model Systems & Compression — Grok-1 structural analysis, routing fidelity, quantization, SAAQ, GGUF/Safetensors tooling.
- Artificial Neuromorphic Supervisor for LLMs — event-driven/SNN state for resource-aware agents; deterministic systems retain safety authority.
- Agoge Model Forge — engineering trajectories → post-training → evaluation → export/serve.
- Theseus Machine Physiology — GPU/CPU/power/VRAM telemetry and machine-state datasets for model/control experiments.
Machine telemetry / physiology
↓
Event-driven / SNN state
↓
Agents + LLM systems
↑
Agent evaluation + engineering trajectories
↑
Model analysis / compression / post-training
- Agent evaluation: measurable coding-agent tasks, trajectories, graders, and failure analysis.
- ML systems: checkpoint tooling, model-format infrastructure, quantization experiments, routing fidelity.
- Post-training: connect curated engineering trajectories to reproducible training/evaluation loops.
- Neuromorphic research: keep SNN work grounded in measurable telemetry/runtime experiments rather than biological imitation claims.
Primary author and maintainer: Raul Montoya Cardenas (rmems).
Recruiter-focused structure and editorial rewrite co-authored on 2026-08-21 with OpenAI ChatGPT — GPT-5.6 Sol. Project-specific AI contributions remain attributed in their respective commits, PRs, experiment records, and release provenance.