About

David Coldeira

Software engineer with 10+ years building production systems, and a physics degree behind the way I build them. Most of my work now is agentic AI: LLMs given real tools, real data and real blast radius, then constrained so they can be trusted in production.

Current Work

Gio — production AI assistant inside a UKAS-accredited lab

Sole developer of GQMLab, the laboratory information management system at Geoquip Marine (ISO/IEC 17025:2017, UKAS #24111). Flask/Python REST API, Flutter PWA, MySQL, Docker, nginx.

Gio is its AI assistant: agentic tool-use over live SQL, charts and report-schema introspection, with defense-in-depth security and role-gated access (admin/engineer/QA/viewer). An accredited lab is an unusually unforgiving place to put an LLM — every answer has to be reproducible and every query auditable.

Marco — domain-agnostic AI decision framework

A Dart/Flutter SDK (~16k lines, 3 packages) with a domain-neutral model — Entity / Goal / Environment / Action / Result — behind a pluggable adapter interface, so A* and genetic-algorithm optimisation, exponential-smoothing prediction and rule-based recommendation are reused across domains. Runs on-device and offline, with an optional local-LLM layer (Ollama / llama.cpp) that explains results in natural language.

SailGenius is the maritime adapter in production: a 39k-line Flutter chartplotter where every LLM-generated waypoint is validated against real water and depth data before it reaches the user.

Froe — local-first agentic coding CLI

A single static Go binary: agent loop, tool implementations (file read/grep/edit/exec/web search/GitHub issues), provider adapters for local backends (llama.cpp, Ollama, LM Studio, vLLM) plus Mistral (EU-hosted) as the cloud fallback — chosen deliberately over a US-based provider for data residency, and a Neovim client over JSON-RPC. What I use daily to build QRL and Bell themselves.

Bell — fine-tuned LLM deployment, live

Qwen3-8B + LoRA fine-tuned on 815 instruction pairs as a self-hosted alternative to a hosted API, with sandboxed Python execution and GDPR-compliant Podman/Caddy deployment on Hetzner (EU). Running on Mistral today. Live at bell.entangledcode.dev.

Independent Research

Quantum Relational Language (QRL)

A quantum programming language where entanglement is a first-class citizen — relations-first, compiling to measurement-based quantum computing. Type system where Switch(d) programs are provably causally nonseparable; 896 passing tests; ~13.6k lines across 31 modules; four backends (Perceval, PennyLane, graphix, Qiskit).

Hardware-validated on a Quandela photonic QPU: S = 2.61 ± 0.08.

Most of the writing on this site is QRL research. It’s here because the standard of proof it demanded is the same one I bring to production AI work.

Background

Education: BSc Physics (Combined Sciences), Heriot-Watt University

Languages: Python, Dart, SQL, JavaScript · Trilingual: English, Spanish, Italian

AI/LLM: Hosted and local LLM APIs, agentic tool-use, retrieval-augmented tool design, LoRA fine-tuning

Stack: Flask, FastAPI, Flutter, PyTorch, Docker, Podman, Linux, PostgreSQL, MySQL, Hetzner, Azure

Contact

I’m open to senior AI engineering and Python backend roles — remote across the EU/UK, or on the ground in Bristol. Happy to talk through any of the systems above in detail.