AI & ML Systems
InfoDetection models with calibrated probabilities, LLM agents with memory and retrieval, knowledge graphs: models built to be right, and to show their working.
PyTorchLLM agentsMCP
I build AI systems for the physical world.
Sensor pipelines, detection models and agents that alert people when conditions change. I built and operated 3DREAMS@SG at NTU, then handed it over to the team in August 2026.
My background is in civil and environmental engineering.
I now build data pipelines, detection models and AI agents. Applied research came first, so I can derive a method, build the software that runs it, and explain both to the people who use the results.
From 2025 to 2026 I designed, built and operated the atmospheric-intelligence platform at the Centre for Climate Change and Environmental Health, NTU: sensor fusion on a ten-minute clock, detection models with calibrated probabilities, and agents that reach out before an episode hits the ground. I handed the platform over to the CCEH team in August 2026, and it runs there today.
Resume
01 — Flagship
Handed over · still running at NTU
3DREAMS@SG
Atmospheric intelligence for Singapore: designed, built and operated by one person. I was the sole designer and software engineer behind the platform at the Centre for Climate Change and Environmental Health, NTU, from 2025 to 2026: four sensor streams fused on a ten-minute clock, two detection stages on every window, and agents that alert before an episode reaches the ground. Handed over in full, the platform still runs today, and was presented by CCEH's founding director at the Regional Action Forum NTU co-hosted with WHO in June 2026. The film below follows one evening of September 2026's haze, from orbit down to the platform's own interface.
4Fused sources
2Detection stages
10yrNEA backfill
10Replay alerts, Feb 2026 episode (v1: 0)
5Geo-Hub viewports
Selected work. Systems in production.
Systems built end to end by one person and run in production: a national-scale haze early-warning platform, agents with real memory, research that keeps its receipts.
Every card below is the system itself: its real data, its real vocabulary. Click through for the full build.
Work you can trust.
Every claim inspectable.
Turned reactive monitoring into early warning: the team is told an episode is likely, with a forecast for when it reaches the surface.
Hourly street-level emission estimates, structurally validated against reference monitoring stations on diurnal shape and peak timing: computer vision wired to a real measurement pipeline.
Peer-reviewed research, presented at international conferences and ministerial events: technical analysis translated into policy.
Six-component detection scoring gated on physics, paired with a cloud classifier that keeps false alarms out of the alert channel.
My tools & your product.
One person, full depth: from the physics of a signal to the pixel that explains it. Every layer below is mine, in production.
How I workWeb & Visualization
InfoNext.js and TypeScript fronts with Three.js and Deck.GL where the data earns it: dashboards that explain themselves instead of decorating.
Next.jsThree.jsDeck.GL
Real-time Data & Backend
InfoSensor ingestion on a ten-minute clock, FastAPI services, PostgreSQL and DuckDB, event-sourced pipelines with vector search: backends that keep their promises.
FastAPIPostgreSQLDuckDB
Cloud & Automation
InfoVercel, Azure and GitHub Actions around everything; serverless and cron where it fits; Teams alerts that reach out before you have to ask, and Teams and Telegram agents that answer on demand.
VercelAzureActions
Choose the pace: embedded partnership or focused sprint.
Engagement model
Embedded / on your team
Full-time includes:
- End-to-end ownership: pipeline → model → interface
- Production discipline: tests, CI, observability
- Comfortable with domain experts & stakeholders
- Based in Singapore, Southeast Asia · remote-first
4–12 weeks / fixed scope
A sprint includes:
- A working system, not a slide deck
- Detection models · agents · data platforms
- Weekly demos, honest status
- Handover with docs your team can run
Questions, answered.
Where are you based, and do you work remotely?
Singapore (SGT), remote-first across Southeast Asia and beyond, with async habits that overlap European mornings and US-West evenings.
What kind of roles are you open to?
Three shapes: AI and systems engineering roles built on sensor and environmental data; solutions and integration roles at weather, climate and geospatial companies; and fixed-scope contract builds of detection systems, agents and data platforms.
What's your core stack?
Python · PyTorch · FastAPI · PostgreSQL and DuckDB on the back; TypeScript · Next.js · Three.js on the front; Vercel, Azure and GitHub Actions around it.
Can I see something running?
Yes: the 3DREAMS@SG case study walks the full system, from sensing and fusion to detection, alerting, and the mission-control UI.
How fast do you respond?
Within a day, usually much faster. Email is best: awvmeijer@gmail.com.
Do you take advisory work?
Selectively: short engagements around detection systems, agent architectures, and environmental data platforms.