San Marcos, Texas · montoyaraul34@gmail.com
Making large models run on hardware you can actually own.
AI engineering student and independent researcher building systems where coding agents can go further with less babysitting — and where model/hardware research stays measurable, reproducible, and safety-bounded.
I coined SAAQ (Spiking Adaptive Activity Quantization) while starting early pipeline work (corinth-canal, Surrogate_Viz.jl): study and compress large MoE models with spiking / neuromorphic ideas on hardware I own (e.g. RTX 5080). I am still learning Julia, Rust, Python, and CUDA. Pursuing AI Engineering @ WGU.
rmems · Limen Neural · GitHub Projects
This portfolio started as an oversized research workspace, then spilled into a noisy org catalog. The work was modularized so each repository has a clear job — but the pieces are meant to compose into one system, not sit as unrelated demos.
The public map is my GitHub Projects board. In short:
Theseus physiology (telemetry + safety governor)
↓
SNN / temporal “nervous system” (Spikenaut + Julia SNN stack)
↓ works alongside / attaches to
Agents + LLMs (Autonomous SE + resource-aware control)
↑
Frontier compression (SAAQ, Grok MoE path) → smaller / spiking-friendly models
↑
Agoge forge (train/eval from engineering trajectories)
Hard invariants I care about more than demo vibes: no silent auto-merge · bounded repair · provenance · measurable gates (CI on pull requests and pushes).
Same five programs as LinkedIn and the Projects tab.
Multi-agent Git worktree execution, engineering-trajectory collection, review/repair loops, and local agent evaluation — so agents ship further with less babysitting. Humans merge.
| worktrees-hives | Early/experimental multi-agent lab in isolated git worktrees; mandatory findings; never auto-merges |
| operation-prometheus | Engineering trajectories (issue→review→patch→validation) for local coding agents |
| NeuralForge-Memory | Experimental RAG / vector DB with MCP |
Open-weight MoE research on real Grok-1 checkpoints: structural dissection, route-preserving quantization / SAAQ, and compression that tracks routing fidelity — not just file size.
| magere-brug | SAAQ lab — recipes, manifests, hybrid MoE/SNN experiments |
| corinth-canal | SAAQ reference pipeline (Rust) |
| xai-dissect | Inspect raw open-weight Grok-1 shards (Rust CLI) |
| grok-ozempic | Grok-scale SNN-style quantization experiments |
| XAIDissect_Viz.jl | Viz of MoE routing / xai-dissect reports |
| hybrid-fusion · engram-parser | MoE→SNN path: hybrid architecture + extract expert “memories” for transfer |
| combine-for-AI | Neutral quant benchmark harness (accuracy, latency, VRAM, routing, …) |
| Surrogate_Viz.jl | Symbolic regression on telemetry (co-origin of the SAAQ term) |
Event-driven / SNN layers that compress machine telemetry into a small internal state so agents can be resource-aware. The goal is an SNN (or event-driven model) that runs alongside an LLM / agent stack — a nervous system, not a full biological imitation. Learned systems propose; deterministic systems protect (hard thermal / resource safety is never overridden by a learned policy). FPGA / HDL work is a deployment path for that nervous system later, not the whole program.
| Spikenaut-SNN | Own SNN from scratch — candidate “nervous system” / transfer target for MoE expert blocks |
| LiquidCortex.jl · SpikeStream.jl · NeuroPulse.jl · TemporalFocus.jl · SynapticDistill.jl | Julia SNN / temporal stack (LSM, streaming features, relevance, spike-coincidence attention, FPGA distillation) |
| silicon-hdl · silicon-bridge | Optional export path to small FPGA targets (learning field, not production HFT silicon) |
| blackwell-kernel-lab | Local GPU path on NVIDIA Blackwell / RTX 5080 for agentic + neuromorphic experiments |
| spike-viz | Visualize SNN encodings and activity |
Telemetry inputs come from Theseus Machine Physiology (below).
Local-first post-training bridge: software-engineering trajectories → LoRA/QLoRA → eval → export/serve.
| agoge-forger | Training forge (PyTorch-first; Rust/JAX options) |
| operation-prometheus | Upstream trajectory datasets that feed the forge |
GPU/CPU/power/VRAM telemetry and machine-state datasets under a Rust safety governor. Mining is one sustained physiology source — not the whole program.
| gaming-telemetry | High-load GPU telemetry for neuromorphic research |
| Theseus-Quarry | Mining / ops telemetry extraction |
| spikenaut-telemetry-etl | Fail-loud cleaning toward published datasets |
| thalamic-relay | Rust hardware orchestration relay (compute telemetry → SNN drive) |
| Ship-of-Theseus-HPC | Workstation notes and hardware diagnostics |
Why modular? (Limen-Neural Consolidation)
Selected repos moved from the Limen-Neural org back under rmems so this reads as a personal research portfolio, not a catch-all product org. The org keeps shared libraries maturing toward publish (neuromod v0.5.1 on crates.io today; others such as nir-rs, brainstem-daemon, myelin-accelerator, axon-encoder still in progress).
Experimental SNN-HFT research hub: Limen-Capital (+ related signal/ledger crates). Not live capital.
Cloud training scaffold (future): Dioscuri-Cloud.
I normally go from issue → pull request, either from a local CLI or through Linear (cloud).
- Agents day to day: Grok Build, Codex, Claude, Cursor, Devin, Kilo, OpenCode, and others — plus shared memory (Ogham / Chroma). Humans merge.
- PR review (rough order): Codex → CodeRabbit → Devin when I want a strong extra pass → Cursor Bugbot · Copilot · CodeAnt · Qodana (local + cloud). I still spot-check with textbooks / local judgment — bots are not ground truth.
- Before open/push: local checks from each repo’s
REVIEW.mdand Qodana CLI when available. - CI: GitHub Actions on pull requests and pushes — early feedback and a rough benchmark of agent (and human) implementations.
- Autonomous SE: worktrees-hives (early/experimental) + operation-prometheus
- Frontier compression / SAAQ: magere-brug + corinth-canal; Grok path via xai-dissect / grok-ozempic
- SNN as nervous system for agents: Spikenaut-SNN + Julia SNN stack; safety stays deterministic
- Physiology inputs: Theseus telemetry path under a safety governor
- Agoge forge: local post-training bridge from SE trajectories
- Education: Pursuing AI Engineering @ WGU
README updated by me (rmems) and with Grok Build: Grok 4.5 (high)