Muriel Demarcus
A lithium battery explosion in a Singapore residential storage shouldn’t be the type of occasion that sometimes sparks a deeptech startup. However for Muriel Demarcus, a seasoned infrastructure danger skilled with three many years of managing billion-dollar tasks throughout Europe and Asia Pacific, it was the second all the things clicked.
“My neighbour’s storage burned to the bottom,” she remembers. “A lithium battery exploded. No person was damage, but it surely was an in depth name, and it stopped me in my tracks. I had spent thirty years managing billion-dollar infrastructure dangers. And right here was a failure mode sitting in a residential storage that no system had caught, as a result of no system was trying.”
Additionally Learn: The AI upskilling wave is actual however the hole it leaves behind is rising
That singular second of frustration led Demarcus to upskill in AI in Singapore, revisit a query she had been asking in management rooms for many years, and ultimately discovered Marsham Edge, a platform constructed round a deceptively easy however deeply tough premise: in high-stakes environments, an warn you can not clarify is an warn you can not act on.
Three brokers, one mission
On the coronary heart of Marsham Edge are three AI brokers Argo, Ken, and Deb — every of which owns a definite layer of the detection pipeline.
Argo manages knowledge ingestion, validation, deduplication, and provenance monitoring. It additionally displays the platform’s personal API endpoints for vulnerabilities — a design resolution drawn from real-world AI system breaches.
Ken runs the detection engine, deciding on and optimising a mannequin stack that features CNN-LSTM for deep sample detection, Random Forest for classification, and a proprietary four-trigger hybrid engine masking statistical envelope, charge density, geometric spike, and physics-informed residual evaluation.
Deb is the coordinating layer: routing duties, assembling findings into structured briefings, and delivering them through dashboard, WhatsApp, or Sign.
“A single-model system provides you a solution,” Demarcus says. “Our agent workforce provides you a course of: safe, detect, and temporary. No black packing containers. Each resolution attributable.”
The multi-agent structure is a deliberate departure from how most AI programs are designed. “Most AI programs are monolithic: one mannequin does all the things. That’s brittle. When the mannequin fails, it fails silently and fully.”
Explainability as structure, not add-on
The phrase “explainability” will get thrown round liberally in AI advertising and marketing. Demarcus has constructed it into the muse of how the system works.
When an alert fires, operators don’t obtain a generic “anomaly detected” flag. They see which of the 4 triggers fired, the precise numerical threshold crossed, the reasoning behind the choice, and the supply knowledge. Within the battery thermal use case, an alert reads one thing like: Set off D fired. Precise thermal charge: 4.2°C/min. Physics mannequin predicted: 2.1°C/min. Residual: 3.2σ. Danger state: Watching temporary (50 per cent). Really useful motion: Cut back load in 45 seconds or vital state predicted.
This method additionally addresses the hallucination drawback that plagues giant language model-based programs in safety-critical contexts. Marsham Edge doesn’t depend on third-party LLM APIs for detection. The detection engine runs on buyer infrastructure, utilizing proprietary fashions grounded in statistical and bodily legal guidelines that structurally get rid of generative ambiguity.
Additionally Learn: Overlook the cloud: Why AI is turning into the brand new heavy business (and what buyers should know)
A two-trigger gate additional reduces false alarms: no single noisy sensor can set off an alert. Two impartial triggers should hearth concurrently earlier than the system points even a Watching Transient.
The battery drawback no person has solved
Certainly one of Marsham Edge’s most compelling use instances is early warning of lithium-ion battery thermal runaway, and Demarcus speaks about it with the urgency of somebody who has witnessed it firsthand.
Thermal runaway is notoriously tough to detect as a result of the failure mode is exponential. By the point a traditional sensor hits its threshold, the response is incessantly irreversible. Most business instruments monitor temperature thresholds and voltage drops, triggers that fireplace too late.
Marsham Edge’s method suits a physics-informed energy-balance mannequin (Newton’s Legislation of Cooling) to every battery’s particular person thermal signature, then repeatedly compares the measured charge of temperature change in opposition to what physics predicts. Validated in opposition to datasets from the Nationwide Renewable Vitality Laboratory (NREL), Sandia Nationwide Laboratories, and NASA, the platform demonstrated early-warning home windows of 220 to 359 seconds forward of normal hardware-level 80°C threshold alarms. “That’s the distinction between a managed intervention and a fireplace,” Demarcus says flatly.
Deployed, validated, and profitable hackathons
Though the startup is lower than a yr outdated, Marsham Edge already has stay deployments. In Could 2026, the complete agent workforce accomplished an integration take a look at in opposition to an artificial OSINT dataset: Argo quarantined all 4 malformed data; Ken detected 18 of 18 marketing campaign posts with zero false positives (F1 = 1.00); Deb delivered a structured analyst briefing in three minutes and 7 seconds.
Shortly after, the platform was deployed on a stay consumer dataset of 174 silica publicity measurements from an underground mining operation in New South Wales, Australia. Ken recognized 31 exceedances — 17.8 per cent of the dataset — with a peak studying of 0.273 mg/m³, or 5.5 occasions the authorized restrict of 0.05 mg/m³. Argo flagged the consumer’s documented use of banned compressed air as an element that elevated their prosecution danger from Class 2 to Class 1.
It’s in opposition to this backdrop that Demarcus received the Epic Hackathon Singapore, competing in opposition to groups she describes as “half my age.”
“Youthful founders typically construct quick and ask questions later. That’s worthwhile. However in safety-critical environments, velocity with out accountability is harmful,” she says. “The hackathon confirmed what I already believed: expertise issues. It teaches you which ones indicators are necessary and that are noise. The brokers deal with the noise. I deal with the accountability.”
Constructing for the regulatory future
Demarcus shouldn’t be merely fixing at this time’s operational issues. She is positioning Marsham Edge on the convergence of three traits she sees as inevitable: mandated explainability below frameworks just like the EU AI Act and Singapore’s AI Confirm programme; the shift to edge and on-premise deployment in regulated industries unwilling to route delicate knowledge by way of third-party clouds; and the broader transfer from monolithic fashions to specialised agentic architectures.
“We’re constructing for the regulatory future, not the regulatory current,” she says.
Additionally Learn: How Asia’s factories are main the best way in industrial AI
The following cease is VivaTech Paris, the place she intends to pursue sovereign cloud companions, defence AI integrators, and buyers who grasp that explainability is quick turning into a compliance requirement slightly than a product differentiator.
“What the worldwide tech neighborhood ought to perceive is that this: Singapore isn’t just a monetary hub. It’s a defence and demanding infrastructure nexus. We’re constructing a platform that solves a common drawback, black-box AI in high-stakes environments, from a rustic that values safety, sovereignty, and belief.”
One yr in, with stay deployments and impartial validation already in hand, Marsham Edge is making a reputable case that the subsequent frontier in AI shouldn’t be uncooked functionality; it’s accountability.
The submit She watched her neighbour’s storage burn down. Now she’s constructing AI that explains itself earlier than catastrophe strikes appeared first on e27.
