Independent applied AI research
The systems behind the intelligence.
Private AI is more than a model on a GPU. This is where I build, test, and study inference, agent behavior, memory, evidence, and the infrastructure that connects them.
Illustrative relationships, not a deployed system blueprint. Internal architecture stays private.
Build history, not live capacity. Ollama supports basic testing and selected embedding workloads.
01 / Explore the work
Independent, AI-assisted research and engineering: what was built, what was measured, and what remains under investigation.
Inference & GPU residency
Retained-weight sleep/wake, model switching, and host-memory contention. Measuring the transition from one model to another, with timing boundaries and failed checks kept in view.
Agentic research & evidence
Self-hosted NVIDIA AI-Q and NeMo workflows. Source tracing, controlled evaluation, citation repair, and replayable evidence across agent handoffs.
Intent & agent systems
A broader intent-layer architecture, explored through early Hermes and Open-WebUI work and a later advisory plugin. Task contracts, memory routing, and handoff boundaries remain distinct from execution authority.
Memory governance
Ogden / Memory Workbench: source provenance, immutable proposals, and separate human decisions. Model-generated proposals do not mutate live memory.
Embedding & retrieval
Nemotron BF16/NVFP4 and Qwen/Nemotron comparisons: serving speed, public retrieval tasks, and source-linked memory. Human-reviewed relevance remains a separate evaluation gate.
61/61
A specific September 2026 pinned-vLLM campaign; 122/122 arithmetic and tool checks. Component validation, not full-system qualification.
196
Validation of a citation-accounting repair and regression comparisons. Software checks, not a research-accuracy score.
2.0x
32 vs. 16 inputs/sec for 8B NVFP4 vs. BF16 in one matched mixed-length, eight-worker soak. Workload-specific, not a general ranking.
55.24GB/s
One four-drive PCIe 5.0 RAID0 array; fio against a prepared 100 GiB file. Not write speed or combined-array throughput; RAID0 has no redundancy.
These records use different workloads, configurations, and timing boundaries. They are not a shared benchmark, live telemetry, commercial deployment claims, or production guarantees. Methods and limitations
02 / Inside the lab
Experiments span physical compute, model serving, agent decisions, retained memory, and evidence. The engineering and the evaluation belong together.
Define the workload, expected behavior, timing boundaries, and failure conditions before a run.
Retain configuration, source identity, observed failures, and the artifacts needed to reproduce a result.
A passing software check does not establish that an AI conclusion is factual.
Keep historical results, bounded experiments, operational deployments, and proposals clearly distinguished.
03 / Engineering notebook
Research briefs from the published record, including experiment scope, observed failures, and remaining questions.
11 research briefs
Four Qwen/Nemotron NVFP4 checkpoints, dataset-specific tradeoffs, and 20,480 requests across bounded mixed-load tests.
Human-authored relevance, blinded grading, and failure-cost gates. Prepared cases are not completed gold labels.
Qwen soak validation, a Super JSON-boundary repair, and an MTPv2 candidate withheld after a CUDA fault.
Qwen3 TTS voice design, reference-based cloning, streaming, and controlled model transitions. Voice-quality review stays separate.
Publisher LoRA integration, concurrent route checks, restart durability, and lifecycle recovery without duplicating the base.
Retained-weight transitions, correctness gates, and the limits of component validation.
A bounded citation-accounting repair, exact-image replay, and focused regression checks.
Source drift, immutable model output, and 19 proposals kept separate from live memory.
Moving from Ollama GGUF to vLLM NVFP4 while retaining 3,045 active memory documents.
Intent contracts, historical regression coverage, failure analysis, and version boundaries.
Native Hermes audio tools, Whisper recognition, Qwen3 TTS, and recoverable deployment.
From infrastructure to intelligence
Independent research. Hands-on engineering. Questions worth exploring.
