Building
intelligent
spatial systems.
I'm Alex Chen — a product engineer who designs and ships AI-native platforms at the convergence of computational intelligence, real-time data infrastructure, and immersive spatial interfaces.
Case Study — Flagship Project
NEXUS
Spatial Analytics
NEXUS transforms enterprise multi-dimensional data into navigable 3D intelligence environments — combining a custom WebGL engine, real-time event pipelines, and large language model reasoning to make complex datasets physically explorable.
The Challenge
Enterprise analytics teams were failing to extract insight from their most valuable datasets — not because the data wasn't there, but because every tool flattened it. 50-dimensional ML feature spaces became 2D bar charts. Network graphs collapsed at scale. Causality got lost in correlation tables.
The Solution
A spatial computing layer that lets analysts navigate data the same way you navigate a city — with depth, directionality, and physicality. Users fly through time-series clusters, manipulate data objects in 3D space, and query in natural language with AI that reasons against spatial context.
Technical Stack
Engineering Highlights
- Custom deferred-shading WebGL renderer with spatial partitioning
- R-tree 3D spatial index adapted for high-cardinality event data
- LLM reasoning layer with scene-context injection for grounded Q&A
- Multi-user collaborative sessions via WebRTC + operational transforms
- Rust/WASM compute module for sub-frame analytics over 10M+ points
Interactive 3D Showcase
The spatial engine,
live and interactive.
Explore the core rendering architecture of NEXUS. Each node represents a data cluster; connections show real-time inference pathways between the AI reasoning layer and the spatial index.
Controls
Live Simulation
Engineering Process
How we made hard architectural decisions.
Discovery & Constraint Mapping
Week 1–2Engineering Decisions
- Ruled out React Three Fiber for the production renderer — abstraction overhead incompatible with 60fps at 10M+ point targets
- Chose ClickHouse over TimescaleDB: columnar storage gives 40× faster aggregation on our query patterns
- Committed to spatial R-tree indexing over grid-based approaches after benchmarking on synthetic enterprise datasets
// R-tree spatial query — 3D bounding box intersection
const query = spatialIndex.search({
minX: view.x - radius, maxX: view.x + radius,
minY: view.y - radius, maxY: view.y + radius,
minZ: view.z - radius, maxZ: view.z + radius,
});
// → returns candidate node IDs in O(log n)Core Renderer Architecture
Week 3–6Engineering Decisions
- Deferred shading pipeline with G-buffer: position, normal, albedo, and emissive channels
- Instanced mesh rendering for up to 100K nodes with zero draw calls per-node
- Custom GLSL shader bank with hot-swap capability for A/B visual testing in production
// Instanced draw call — 100K nodes, 1 draw call
renderer.drawInstanced(
nodeGeometry,
holographicMaterial,
instanceBuffer, // Float32Array: mat4 per node
visibleNodeCount, // frustum-culled count
);AI Reasoning Integration
Week 7–10Engineering Decisions
- Spatial context injection: serialize visible scene graph as structured JSON into LLM system prompt
- Streaming response with partial rendering: each reasoning token triggers incremental scene annotation
- Caching layer with semantic similarity matching to avoid redundant inference on repeated queries
// Spatial context → LLM system prompt injection
const context = sceneGraph.serialize({
nodes: visibleNodes.slice(0, 200),
topology: 'adjacency-list',
format: 'compressed-json',
});
const stream = await openai.chat.completions.create({
model: 'gpt-4o',
messages: [{ role: 'system', content: SYSTEM + context }],
stream: true,
});Collaborative Real-time Layer
Week 11–13Engineering Decisions
- Operational transformation chosen over CRDT for scene mutations — lower memory footprint for our data shape
- WebRTC mesh for ≤8 users; SFU relay kicks in above that threshold to prevent O(n²) bandwidth scaling
- Presence cursors rendered as separate instanced geometry pass — invisible to frame budget
// OT document mutation — concurrent safe
const op = SceneOT.transform(
localOp,
remoteOp,
serverRevision
);
scene.apply(op);
broadcast({ type: 'op', op, rev: ++localRev });Platform Capabilities
Six capabilities that make NEXUS irreplaceable.
Each feature is deeply integrated — not bolted on. Click any card to expand the engineering detail.
Spatial Navigation
Fly through data like a physical environment.
Natural Language Queries
Ask questions about what you see in 3D space.
Real-time Streaming
10M+ events/second. Zero lag. Zero sampling.
Multiplayer Sessions
Share a spatial view. Work on it together, live.
Anomaly Detection
94.7% accurate. Highlights itself in the scene.
Spatial Report Export
Export an annotated PDF of any 3D view state.
Impact & Outcomes
Numbers that speak without a deck.
ARR Trajectory
Months 1–14 post-launch
Deployments by Region
200+ enterprises
"NEXUS changed how our data science team communicates findings to leadership. What used to be a 20-slide deck is now a 10-minute spatial walkthrough. Decision velocity tripled."
"I've used every major BI tool. NEXUS is the first one that made our fraud detection dataset feel like a physical place I can explore. We caught a pattern in 2 hours that had evaded automated systems for 6 months."
About
Alex Chen — product engineer
and spatial systems builder.
I spend my time at the boundary between software engineering and product design — building systems that are both technically rigorous and genuinely delightful to use. I believe the best products feel inevitable in retrospect.
My background spans graphics programming, distributed systems, and applied AI. I built my first Three.js scene in 2015 before it was fashionable; I've been chasing the feeling of "what if data felt like a place?" ever since.
NEXUS is the culmination of that question — a platform that treats enterprise data not as a spreadsheet, but as a navigable, collaborative, AI-augmented world.
Open Source
Craft Profile
Career Timeline
Sole architect and primary engineer. Designed the rendering pipeline, data infrastructure, AI integration, and collaborative layer from scratch.
Foundry graph visualisation layer and data ontology pipelines. Shipped features used by hundreds of Fortune 500 analysts.
Built real-time fraud detection dashboards and internal analytics infrastructure. Introduced Three.js-based 3D transaction graph as an internal prototype.
Let's build something impossible.
I'm selectively available for ambitious projects — AI-native products, spatial interfaces, high-performance web applications, or strategic engineering advisory. If you're building something that shouldn't be possible yet, I want to hear about it.