Quantum + AI · intelligent instruments, LLM fine-tuning, quantum-classical hybrids

Physics and NMR background, sixteen years teaching quantum mechanics, now building scientific instruments where the sensor, the firmware, the signal pipeline, and the model live in one project — and, lately, small quantum subroutines that sit behind stable classical interfaces in an LLM stack.

Three open research projects asking where a small quantum routine can do real work inside an LLM pipeline. Each ships a tested library, a live demo Space, and honest benchmarks. These are research prototypes on quantum simulators (QVerify also runs on real IBM hardware); no quantum advantage is claimed — the goal is working, honestly measured quantum subroutines behind classical interfaces.

Translates each LLM reasoning step to propositional CNF, grounds it over a finite universe, and checks logical consistency with Grover's algorithm — on a CPU simulator or IBM's 156-qubit Heron r2 processor.

v1.0.1 · 485 tests · 14 verified IBM Heron r2 hardware runs.

Picks the best subset of k tools from N with the Quantum Approximate Optimization Algorithm, accounting for the tool-pair interactions a greedy ranking ignores.

v0.2 · approximation ratio 0.915 at N=16 (vs greedy 0.735).

A variational quantum circuit that routes tokens to experts in a Mixture-of-Experts layer: n qubits give 2^n basis states, one per expert, so a tiny circuit scores an exponential number of experts.

Phase 1 prototype (toy MoE) plus a full-scale 128-expert / 7-qubit design.

Hardware-aware AI. I'm most useful when a problem needs someone who can design a PCB, flash a microcontroller, tune a neural network, and ship the whole thing as a product. Current focus areas:

• Scientific instrumentation with integrated AI (microscopy, radio astronomy, sensor arrays)
• Quantum-classical hybrids: small quantum subroutines inside LLM pipelines
• Multi-GPU inference infrastructure for local LLM stacks
• Computer vision for biological and medical signals
• Language model efficiency (quantization, parameter-constrained training)

sovereign-ai-stack. Self-hosted local AI stack on a 3-GPU workstation with Ollama, vLLM, API key routing, GPU mutex, and a noVNC web UI. For people who want to own their models and their data.

parameter-golf. Personal case-study of my participation in the OpenAI Model Craft Challenge. Two submissions, 1.2421 and 1.1431 bits per byte on FineWeb, within 0.003 of the March 20 leaderboard top.

AI-Microscope. Computer vision and ML for optical microscopy: autofocus metrics, illumination analysis, specimen classification. Runs a ToupCam camera through a FastAPI backend with a React viewer and real-time WebSocket streaming.

HydrogenEye. Radio astronomy receiver built from a 28-euro RTL-SDR dongle. Detects the 21 cm hydrogen line with roughly 35 dB SNR. Raspberry Pi Zero W + Android app over Wi-Fi.

BioRNG. True random number generation from live fish behavior via computer vision. Entropy validated against the NIST SP 800-90B battery.

VRAM-Pressure-Scheduling. VRAM-aware GPU scheduling for multi-GPU AI workstations: priority preemption, NVLink topology constraints, thermal-aware placement. Six months of production results included.

api-tester-app. Mobile tool for testing and managing API keys across AI inference providers. Works as a PWA and as an Android APK.

AI-Gas-Analyzer. Neural classification of gas mixtures from metal-oxide sensor arrays. Inference on microcontroller hardware.

SO-101-LeRobot-Industrial-Swarm. Multi-arm robotic swarm coordination through a WebSocket bus, built on the Hugging Face LeRobot SO-101 arm.

1420MHz-Feed-Horn. Waveguide feed horn design for 1420 MHz hydrogen line radio astronomy. Pairs with HydrogenEye.

Python, C, C++, Rust when needed, TypeScript for UI, FastAPI and Flask for backends, React for frontends, PyTorch and ONNX for models, Docker for deployment, Linux everywhere. For the quantum work: PennyLane, Qiskit, IBM Quantum, Gradio, Hugging Face, and GitHub Actions for CI and auto-deploy.

• GitHub organization for the quantum work: Quantum-Labor
• Hugging Face: Laborator
• This profile: github.com/SergheiBrinza