Engineering notes · Field evidence
Ideas tested against real systems.
Technical deep-dives, implementation notes and reproducible methods from building AI infrastructure, Rust accelerators and open-source developer tools.
Read the work behind the work.
Methods, trade-offs and limitations travel with the implementation. Claims remain tied to the linked source and version.
Cost Control for Production AI Agents
Where the money goes in an agent run, how to attribute spend per task rather than per month, and which of caching, batching and budgets to build first.
Database Access Patterns for AI Agents
Four ways to give an agent access to a production database, what each one actually defends against, and why the audit trail matters more than the query language.
PyO3 in Production: What We Learned
Packaging, ABI choices, GIL handling, error mapping and monkeypatch safety: the decisions behind shipping Rust extensions to Python users.
The GIL Ceiling in Python AI Frameworks
Why LiteLLM, LangGraph and CrewAI stop scaling before the network does, what the GIL actually serialises, and how to tell a GIL wall from an I/O wall.
Notes from Shipping 28 Projects in a Year
What we learned shipping 28 MIT-licensed projects: agent runtimes, Rust accelerators, dev tools and governance, and which themes actually turned out to hold.
Drop-in Compatibility as a Design Principle
Why most of the Neul Labs portfolio refuses to change your config, your CLI or your imports: the rule, the trade-offs, and the cases where we break it.
The Agentic Infrastructure Stack
A reference architecture for an agentic stack: workspace runtime, memory, ORM policy, orchestration and payments, built from six Neul Labs primitives.
Why Rust Accelerates Python AI Infrastructure
The GIL and single-threaded serialization bottleneck production AI agents. Rust with PyO3 gives drop-in replacements for LiteLLM, LangGraph and CrewAI.
Making LangGraph Checkpoints 700x Faster
Checkpoint serialization dominates production LangGraph runs. fast-langgraph uses Rust to cut write time 737x, read time 151x and end-to-end time 2.8x.
Running AI Coding Agents in Parallel
brat is an open-source multi-agent harness that coordinates Claude Code, Aider and Codex on one codebase with crash-safe state, merge queues and a dashboard.
Python Task Queue Without Redis
fastworker is a brokerless Python task queue that removes Redis and RabbitMQ. Priority queues, auto-discovery, distributed workers and a built-in dashboard.