Labs
Research-grade systems experimentation — validating AI infrastructure, benchmarking inference and retrieval, studying human-AI interaction, and publishing reproducible technical findings. Not demos. Not hobby projects. Experimental work with methodological rigor.
All reports
Running a 70B model on Apple Silicon, end to end
A reproducible local inference setup with quantization, KV-cache tuning, and honest token-throughput numbers.
Evaluating RAG retrieval with adversarial questions
What happens to answer quality when the right chunk is present but buried under near-duplicates.
Five failure modes in production agent loops
An observational study of 200 real agent runs — cataloguing the repeating failure patterns that kill multi-step tasks before the finish line.
Lab stack
Tools & technologies
Ollama
Local model serving
llama.cpp
CPU/GPU inference
Apple Silicon (M-series)
On-device GPU acceleration
Hugging Face
Model hub & transformers
LangChain / LlamaIndex
LLM orchestration
ChromaDB
Open-source vector store
pgvector
Postgres vector search
Python + FastAPI
Eval pipelines & inference APIs
What we experiment on
Every report is fully reproducible — we publish the hardware specs, software versions, prompts, and evaluation criteria used. If a number seems too good, we explain why.
- Local inference on consumer hardware
- Quantization tradeoffs (quality vs. speed)
- RAG retrieval evaluation
- Embedding model comparisons
- Agent architecture patterns
- Apple Silicon GPU benchmarks
Stay close to the work
New lab reports in The Signal
Every new experiment lands in The Signal before it's announced anywhere else.
This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.