Muriel Demarcus
A lithium battery explosion in a Singapore residential garage is not the kind of event that typically sparks a deeptech startup. But for Muriel Demarcus, a seasoned infrastructure risk professional with three decades of managing billion-dollar projects across Europe and Asia Pacific, it was the moment everything clicked.
“My neighbour’s garage burned to the ground,” she recalls. “A lithium battery exploded. Nobody was hurt, but it was a close call, and it stopped me in my tracks. I had spent thirty years managing billion-dollar infrastructure risks. And here was a failure mode sitting in a residential garage that no system had caught, because no system was looking.”
Also Read: The AI upskilling wave is real but the gap it leaves behind is growing
That singular moment of frustration led Demarcus to upskill in AI in Singapore, revisit a question she had been asking in control rooms for decades, and eventually found Marsham Edge, a platform built around a deceptively simple but deeply difficult premise: in high-stakes environments, an alert you cannot explain is an alert you cannot act on.
Three agents, one mission
At the heart of Marsham Edge are three AI agents Argo, Ken, and Deb — each of which owns a distinct layer of the detection pipeline.
Argo manages data ingestion, validation, deduplication, and provenance tracking. It also monitors the platform’s own API endpoints for vulnerabilities — a design decision drawn from real-world AI system breaches.
Ken runs the detection engine, selecting and optimising a model stack that includes CNN-LSTM for deep pattern detection, Random Forest for classification, and a proprietary four-trigger hybrid engine covering statistical envelope, rate density, geometric spike, and physics-informed residual analysis.
Deb is the coordinating layer: routing tasks, assembling findings into structured briefings, and delivering them via dashboard, WhatsApp, or Signal.
“A single-model system gives you an answer,” Demarcus says. “Our agent team gives you a process: secure, detect, and brief. No black boxes. Every decision attributable.”
The multi-agent architecture is a deliberate departure from how most AI systems are designed. “Most AI systems are monolithic: one model does everything. That is brittle. When the model fails, it fails silently and completely.”
Explainability as architecture, not add-on
The word “explainability” gets thrown around liberally in AI marketing. Demarcus has built it into the foundation of how the system works.
When an alert fires, operators do not receive a generic “anomaly detected” flag. They see which of the four triggers fired, the exact numerical threshold crossed, the reasoning behind the decision, and the source data. In the battery thermal use case, an alert reads something like: Trigger D fired. Actual thermal rate: 4.2°C/min. Physics model predicted: 2.1°C/min. Residual: 3.2σ. Risk state: Watching brief (50 per cent). Recommended action: Reduce load in 45 seconds or critical state predicted.
This approach also addresses the hallucination problem that plagues large language model-based systems in safety-critical contexts. Marsham Edge does not rely on third-party LLM APIs for detection. The detection engine runs on customer infrastructure, using proprietary models grounded in statistical and physical laws that structurally eliminate generative ambiguity.
A two-trigger gate further reduces false alarms: no single noisy sensor can trigger an alert. Two independent triggers must fire simultaneously before the system issues even a Watching Brief.
The battery problem nobody has solved
One of Marsham Edge’s most compelling use cases is early warning of lithium-ion battery thermal runaway, and Demarcus speaks about it with the urgency of someone who has witnessed it firsthand.
Thermal runaway is notoriously difficult to detect because the failure mode is exponential. By the time a conventional sensor hits its threshold, the reaction is frequently irreversible. Most industry tools monitor temperature thresholds and voltage drops, triggers that fire too late.
Marsham Edge’s approach fits a physics-informed energy-balance model (Newton’s Law of Cooling) to each battery’s individual thermal signature, then continuously compares the measured rate of temperature change against what physics predicts. Validated against datasets from the National Renewable Energy Laboratory (NREL), Sandia National Laboratories, and NASA, the platform demonstrated early-warning windows of 220 to 359 seconds ahead of standard hardware-level 80°C threshold alarms. “That is the difference between a controlled intervention and a fire,” Demarcus says flatly.
Deployed, validated, and winning hackathons
Although the startup is less than a year old, Marsham Edge already has live deployments. In May 2026, the full agent team completed an integration test against a synthetic OSINT dataset: Argo quarantined all four malformed records; Ken detected 18 of 18 campaign posts with zero false positives (F1 = 1.00); Deb delivered a structured analyst briefing in three minutes and seven seconds.
Shortly after, the platform was deployed on a live client dataset of 174 silica exposure measurements from an underground mining operation in New South Wales, Australia. Ken identified 31 exceedances — 17.8 per cent of the dataset — with a peak reading of 0.273 mg/m³, or 5.5 times the legal limit of 0.05 mg/m³. Argo flagged the client’s documented use of banned compressed air as a factor that elevated their prosecution risk from Category 2 to Category 1.
It is against this backdrop that Demarcus won the Epic Hackathon Singapore, competing against teams she describes as “half my age.”
“Younger founders often build fast and ask questions later. That is valuable. But in safety-critical environments, speed without accountability is dangerous,” she says. “The hackathon confirmed what I already believed: experience matters. It teaches you which signals are important and which are noise. The agents handle the noise. I handle the accountability.”
Building for the regulatory future
Demarcus is not merely solving today’s operational problems. She is positioning Marsham Edge at the convergence of three trends she sees as inevitable: mandated explainability under frameworks like the EU AI Act and Singapore’s AI Verify programme; the shift to edge and on-premise deployment in regulated industries unwilling to route sensitive data through third-party clouds; and the broader move from monolithic models to specialised agentic architectures.
“We are building for the regulatory future, not the regulatory present,” she says.
Also Read: How Asia’s factories are leading the way in industrial AI
The next stop is VivaTech Paris, where she intends to pursue sovereign cloud partners, defence AI integrators, and investors who grasp that explainability is fast becoming a compliance requirement rather than a product differentiator.
“What the global tech community should understand is this: Singapore is not just a financial hub. It is a defence and critical infrastructure nexus. We are building a platform that solves a universal problem, black-box AI in high-stakes environments, from a country that values security, sovereignty, and trust.”
One year in, with live deployments and independent validation already in hand, Marsham Edge is making a credible case that the next frontier in AI is not raw capability; it is accountability.
The post She watched her neighbour’s garage burn down. Now she’s building AI that explains itself before disaster strikes appeared first on e27.