A quantum language, an assistant that runs it live, and the agent that builds both.
Bristol, UK · EU citizen · ten years in software · physics graduate · Linux (Debian) user
I design, build and run production systems from scratch: architecture, backend, frontend, deployment. When a language model goes into one of them, deterministic code checks its output before it reaches a user or a database, and permissions are enforced by the system rather than by the prompt.
QRL — a relations-first quantum
programming language where causal structure is a type-level primitive: a
program that type-checks as Switch(d) is provably causally nonseparable, a
result no existing quantum programming language can state. Full compiler
pipeline to MBQC measurement patterns across four backends (Perceval,
PennyLane, graphix, Qiskit). Hardware-verified on Quandela's photonic QPU
(S = 2.61 ± 0.08). ~13,600 lines · 896 tests · 31 modules ·
Zenodo preprint ·
QRL Book
Bell — the fine-tuned assistant that runs QRL live. Plain English in, real QRL code executed, real numbers out — not the model guessing a number. Provider-swappable: running on Mistral (EU-hosted) today, a self-hosted Qwen3-8B fine-tune (815 training pairs) in standby. Sandboxed execution, Podman/Caddy on Hetzner. Try it.
Froe — the local-first coding agent used to build both. A single static Go binary for the terminal and Neovim. Source never has to leave the machine, the LAN, or — for anyone working under UK/EU data-residency rules — the EU. MIT licensed, public.
Gio — an AI assistant over a live laboratory database · day job, Geoquip Marine
LLM tool-use inside GQMLab, the LIMS I built
for a UKAS-accredited geotechnical laboratory. Natural-language SQL, charts and live
report-schema introspection on a read-only database user with query validation and
role-gated access. In production for the lab's engineers.
Stack: Python/Flask · Flutter · MySQL · Docker · nginx · Hetzner
Marco — an LLM-powered passage planner · Forza7 sailing suite
Reasons over live weather, tide and boat-performance data to propose offshore
routes. Every AI-generated waypoint is validated against real water and depth data
before the sailor sees it. Flutter apps on a FastAPI/PostgreSQL backend.
- Putting an LLM on a Production Database, Read-Only by Construction Four independent layers, only one of which actually matters, and why the prompt is the layer you should trust least.
- Don't Let the Model Do the Physics Keep the model to reasoning and let deterministic code own every number that reaches a user.
More at dcoldeira.github.io.
david@entangledcode.dev · entangledcode.dev · linkedin.com/in/dcoldeira