AI Systems Portfolio on Curtis Nix

Building an eval framework after my benchmark flipped

Thu, 09 Jul 2026 00:00:00 +0000

My June inference bench said llama.cpp was faster than oMLX for specific models. Three weeks later a rerun said oMLX was faster by 46 percent. After investigating in circles, I ditched my own benchmarking scripts. I set up a self-hosted LangFuse instance on my home Kubernetes cluster, centered my benchmarking process around this, and ran the previous bench through it as a sanity check.

llama.cpp vs oMLX on M1

Wed, 08 Jul 2026 00:00:00 +0000

Two 4-bit inference benches on a 32 GB M1. A 5-arm speculative-decoding comparison on Qwen3-8B (plain decode won every matchup) and a 2-arm oMLX vs llama.cpp sweep across four daily-driver models (oMLX faster on three, wash on one).

MLX vs llama.cpp on K8s triage (bench v2)

Thu, 25 Jun 2026 00:00:00 +0000

Follow-up eval comparing llama.cpp and MLX on a focused K8s troubleshooting workload. 13 model+backend combinations, 20 triage scenarios, LLM-as-judge scoring. Stay on llama.cpp confirmed; granite-4.1-8b-tuned emerged as the best speed-at-quality default.

Frontier vs local LLMs on structured HTML

Sat, 20 Jun 2026 00:00:00 +0000

22-prompt structured-generation benchmark across 17 frontier and 5 local LLMs on HyperFrames HTML composition. Deterministic grading via lint and render. kimi-k2.5 scores a perfect 22/22 at 36.0s median; best local gemma3-12b at 14/22. The real frontier advantage is consistency in the tail, not raw accuracy on routine work.

ASR engines on Apple Silicon

Mon, 15 Jun 2026 00:00:00 +0000

Four-engine reproducible benchmark of whisperx, parakeet-mlx, mlx-whisper, and whisper.cpp on Apple Silicon. 8 cohorts across 16 LibriSpeech clips, scored on WER and real-time factor. parakeet-mlx tdt-1.1b wins on the speed-at-quality frontier; mlx-whisper large-v3 disqualified itself on a hallucination loop.

Zero-shot voice cloning on Apple Silicon

Wed, 10 Jun 2026 00:00:00 +0000

Replaced a planned Piper fine-tune with zero-shot F5-TTS-MLX for video narration on Apple Silicon. A 12.24-second reference clip and a tuned ODE solver produce usable voice clones at ~19x real-time, setting the pipeline’s latency floor and shaping worker-daemon concurrency.

Where a video pipeline spends its time

Fri, 05 Jun 2026 00:00:00 +0000

End-to-end timing benchmark of the 6-step Windmill video-creation pipeline on Apple Silicon. 16 clips x 3 reps, 48 runs, RTF 1.0 versus RTF 0.055 for parakeet-mlx in isolation; the ~18x production tax is ffmpeg pre/post-processing, not Windmill scheduling.

OTA firmware over ESP-NOW

Mon, 01 Jun 2026 00:00:00 +0000

A custom four-message OTA protocol for pushing firmware between ESP32s over ESP-NOW (no WiFi, no BLE). Protocol and partition layout proven end-to-end on small payloads; the full 742 KB push reaches 97% before crashing on memory pressure.

Two-board ESP32 lab reachable by chat

Mon, 25 May 2026 00:00:00 +0000

Bootstrap of a permanent two-board ESP32 lab with an always-on agent reachable via chat. Validated end-to-end with 1 Hz ESP-NOW pingpong (100/100 frames) and RSSI-based distance estimation.

MLX vs llama.cpp on M1 Pro (bench v1)

Fri, 15 May 2026 00:00:00 +0000

Side-by-side evaluation of llama.cpp and MLX (via oMLX) for self-hosted inference on a single M1 Pro. Custom blind-scoring harness, 6 matched model pairs, 780 measurement rows. Zero migration recommended, and a silent audio bug surfaced in the MLX path.

Autonomous Infrastructure Anomaly Detector (AIOps)

Mon, 01 Jan 0001 00:00:00 +0000

What: A predictive operations model that ingests live or historical server/cluster metrics to classify system health, proactively warning of node failures or memory leaks before orchestration panics.

Why: Bridges AI/ML coursework with existing infra skills (Talos, Debian, K8s). The dataset and problem are straightforward enough for a WGU project but real enough to be useful.

Approach: Python + Scikit-Learn classification pipeline. Historical metrics → binary classifier (normal vs. anomalous). Hyperparameter tuning, ensemble methods, regularization.

Local LLM Performance & Throughput Predictor

Mon, 01 Jan 0001 00:00:00 +0000

What: A regression model trained on self-benchmarked hardware data that predicts tokens/second or VRAM consumption from inference parameters.

Why: Uniquely positioned. Already run local LLMs (Ollama, MLX) and have the hardware to generate real training data. The dataset is self-created, which strengthens the D683 proposal.

Approach: Python + Scikit-Learn regression. Scripted benchmarking loop across quantization levels, model sizes, context windows, and prompt lengths. Output: continuous TPS or TTFT prediction.

Data:

Smart Greenhouse & Hydroponic Climate Optimizer

Mon, 01 Jan 0001 00:00:00 +0000

What: An agricultural optimization model that parses environmental and nutritional vectors to predict crop growth health or determine optimal feeding schedules.

Why: Intersects with the aquaponics interest already on the radar (housekeeping tracker: “strongest energy thread is electronics + plants/aquaponics”). Turns a genuine hobby into coursework.

Approach: Python + Scikit-Learn. Tabular CSV data with environmental sensor readings. Either regression (predict growth rate) or classification (stress status).

Data:

Format: Tabular .csv with cultivation metrics
Features: Ambient temp, relative humidity, photoperiod hours, water EC/nutrient level, pH
Target: Continuous (biomass/growth rate mm) or categorical (stress T/F)
Cleanliness: Avoid raw images, tabular only. UCI Crop Recommendation dataset works out of the box.

Fit: D683 (custom dataset). UCI dataset is pre-cleaned, well-documented, and instantly approvable.