🤓 Latest Submissions
Quantum Dot Live Twin
Quantum dot labs read device states via dispersive or RF measurements, operating much faster than the classification loops meant to track them. When a twin lags behind, drift can go unnoticed until valuable fridge time is lost recharacterising the device. We built a live digital twin designed to stay in sync with a quantum dot’s changing state under exactly that kind of pressure. It's a monitoring and synchronisation layer a lab can trust between tuning runs, not a system that wrestles the operator for gate control. The frames come from a real QArray-simulated device trajectory, powered by Rust for speed. In a 2000-frame run on an AMD MI300X (PyTorch/ROCm), the serial CPU baseline hits a worst-case lag of 0.155s. GPU batching without load-shedding actually performs worse under pressure, spiking to 3.304s as backlog builds toward the end. A rule-based triage agent solves this, holding worst-case lag to just 0.021s while cutting compute further, routing 235 full passes, 15 light passes, and 44 skips in real time based on queue depth, staleness, and drift signals, a decision actively used by the control loop. An LLM supervisor (Fireworks AI) sits above the triage agent, monitoring backlog trends in the background to retune thresholds without slowing the real-time path. The adaptive, LLM-tuned threshold architecture is running end-to-end. Next up is longer soak testing to capture threshold updates under sustained drift, which is a clear path to differentiation as the system scales to real hardware. The same stream is exposed as a true QCoDeS Instrument, so a lab can connect this twin directly to their instruments with no extra integration layer. That’s the difference between a demo and something a lab could plug in this week. Results arrive in an interactive control console built for the bench, not a slide deck. Future work includes moving from synchronous micro-batching to a fully asynchronous pipeline and training the classifier weights.
13 Jul 2026
SimQuantum: Agentic AI for Quantum Chips
Right now, somewhere in a quantum devices lab, a PhD student is spending hours cooling down a quantum chip, then carefully adjusting voltage knobs until exactly one electron sits in each tiny semiconductor pocket called a quantum dot. Miss the mark even slightly, and the whole setup fails. This painstaking process, known as device tuning, is one of the biggest hurdles to scaling quantum computers. SimQuantum fixes that. It explores a 2D voltage space using a 6-stage decision loop modeled as a POMDP, the same framework used in robotics when systems have to act without perfect information. At each step, it measures voltage, checks if the reading makes physical sense (Data Quality Control), classifies what it sees with a 5-model CNN ensemble trained on 51,000 simulated charge stability diagrams, updates a particle filter belief over possible charge states, and picks the next voltage move using Bayesian optimisation. A safety critic enforces strict voltage limits, and every choice is logged for audit purposes. On top of that, an LLM called Dr. Q (Qwen3-8B) serves as a live copilot in the interface. It sees the full agent state at all times: current stage, voltage position, CNN confidence, belief probability of the target state, and anomaly flags. Users can ask it anything in plain language, and it explains the agent’s actions, flags issues, and produces a summary at the end. The full stack was built on AMD MI300X via ROCm, with Qwen2.5-1.5B served via vLLM as the original inference target. The Streamlit demo runs on HuggingFace Spaces with Dr. Q connected via Fireworks AI (Qwen3-8B) as a portable fallback, pointing the sidebar URL back at an MI300X endpoint switches the backend instantly seamlessly. The agent is able to scan a large 2‑D voltage space (16 V²), automatically detect where a charge transition happens, and correctly label the charge state, all on it's own. Next goal, Navigation: getting the agent to move precisely to the target point on purpose.
10 May 2026
Infrastructure Investment Intelligence Agent
Our solution tackles a key challenge for telecom operators and infrastructure investors across the SADC region, where decisions about where to deploy connectivity infrastructure are hindered by sparse, outdated, or fragmented data. Traditional approaches rely on costly field surveys and incomplete reports, often leading to inefficient capital allocation and missed opportunities to reach underserved communities. We built an intelligent agent system powered by IBM watsonx Orchestrate that synthesises tower coverage data, settlement locations, and socioeconomic indicators to generate actionable investment insights. It brings together five tools: a readiness index ranking countries by investment potential, a coverage gap analyser detecting settlements beyond tower reach, a deployment cost estimator reflecting local logistics, a multi-country comparison tool, and a settlement prioritisation engine that uses geographic heuristics when demographic data is missing. A key strength of our approach is transparent data handling. Instead of hiding data gaps, the system uses proxy metrics such as settlement clustering, proximity to infrastructure, and strategic location to infer density and feasibility while clearly signalling confidence levels. Accessible via web and WhatsApp, the agent enables field teams to query coverage gaps, compare markets, and estimate deployment costs from anywhere, reducing reliance on specialised GIS staff. The solution supports mobile operators expanding rural coverage, development organisations targeting digital inclusion, and investors assessing SADC opportunities. By lowering the time and cost of analysis while improving decision quality, it accelerates connectivity deployment across Southern Africa and demonstrates the power of agentic AI in complex, data-limited environments.
23 Nov 2025
