Project 03 · 2023
AI/ML Leak Detection & Pipeline Validation
IntelliFlux Controls · UX Engineer
A map-first tool that turned illegible plant data into a shared language — for field engineers running validations and for the C-suite closing deals. Built inside IntelliFlux, it outlived the company: it's still running client sites across the US and Asia today under IntelliFlux's successor, EFAS Technology.
Interactive prototype
Open prototypeBefore
Nothing anyone could read There was no map. No spatial tool at all. The only window into plant health was a wall of Grafana graphs — and outside the chemical engineers and data scientists who lived in that data daily, almost no one could tell what those graphs were actually saying. Not the field operators. Not sales. Not the C-suite standing in front of a client. The data existed; the understanding didn't.
The Problem
Six ways the gap showed up.
Illegible by design
Grafana graphs are built for people who already know what they're looking for. Everyone else — including the people paying for the plant — was locked out.
Two worlds, no bridge
Simulation results came back as raw CSVs. The network lived on a map. Correlating them meant manual, eyeball cross-referencing every time.
Days, not hours
With no shared surface for model output and field reality, validating a hydraulic model was a slow reconciliation chore, not a workflow.
Technical data, buried
Elevation, roughness, Reynolds number, flow rate — the parameters engineers needed existed, but not in a form anyone could act on mid-investigation.
No spatial trace
No way to isolate a node range, see what was upstream, or see what a leak would take down with it. Every investigation started from zero.
Two audiences, one tool
Field operators needed granular detail. The C-suite needed to walk a client through the same data and land a business case. It had to do both.
The Research
Wider than a typical engineering tool ever needs to go. I ran research across a wider set of stakeholders than most engineering tools ever touch. On the technical side: data scientists, chemical engineers, data analysts, and PLC engineers doing the actual validation work. On the business side: several of the company's actual clients, plus IntelliFlux's own C-suite, marketing, and sales teams — because this tool had to hold up in a sales conversation, not just a control room. The pattern on the technical side was consistent: the gap wasn't a missing feature, it was a missing bridge between simulation data and the map. On the business side, the ask was related but distinct — the same data needed to read as proof, not just as information.
The Journey
From a clickable map to a full validation tool. We didn't start with the finished product. The first version was deliberately small: a static HTML map embedded in the portal, with clickable link points marking each asset — barely more than a wayfinding layer over a picture. It was enough to prove the core bet: that people who couldn't parse a Grafana graph could immediately understand a map. From there, each iteration added a real layer of capability — flow-based visualization, live metadata, CSV-driven validation — until that static map became the full interactive tool. I worked directly with the VP of Technology to pin down requirements, then led implementation with the dev team through iteration, from early prototypes tested in Felt through to the integrated version shipped inside Apricot.
Early Iterations
The first prototypes tested the core idea at low fidelity: could a map make a complex pipeline network easier to read, navigate, and investigate than a wall of graphs? We built the next round in Felt — a quick, collaborative mapping tool — to turn the concept into a dynamic mockup we could actually explore and analyze with stakeholders.
The Solution
An elemental solution that carries massive power of significance. The fix wasn't a new algorithm. It was refusing to let the map and the data stay strangers — and building it clean enough that it could double as a client-facing demo.
- Interactive map → pipelines rendered with flow-based color gradients; junctions and nodes marked clearly enough to scan at a glance
- Metadata panel → auto-updates with the technical parameters that mattered (elevation, roughness, Reynolds number, flow rate) the moment you touched a node
- Layer controls → toggle pipes, junctions, and sub-networks so a 500-node system didn't have to be read all at once
- CSV import → model-to-field validation with visual overlays, so a simulation result stopped being a spreadsheet and started being a picture
- Leak detection workflow → operators could isolate node ranges, apply filters, and see anomalies in the context of the network around them
Launched on Apricot Portal
This version shipped on IntelliFlux's Apricot Portal and was rolled out to major client sites. The same map-first presentation that made validation faster for engineers also became the client-side view — turning complex leak data into something clear enough to stand in front of.
The Outcome
Days → Hours
Validation cycle time
$ Millions
Saved for clients
US + Asia
Still serving today
The idea outlasted the company that built it. After IntelliFlux wound down, the tool carried forward under its new CEO as EFAS Technology — still serving clients across the US and Asia today, and still giving their C-suite a concrete, visual case to make in front of the people writing the checks.
Continuing the Exploration
Manifold The core insight from this project — that leak detection is really a spatial reasoning problem, not a data problem — was worth testing again outside any proprietary system. So I rebuilt the interaction model from scratch as Manifold, a standalone, independent prototype.
Manifold uses none of IntelliFlux's or EFAS's code, data, or assets — it's a personal reinterpretation of the same problem, built to keep pressure-testing the idea.
- Schematic-first, not GIS-dependent → renders a P&ID-style schematic instead of requiring real geographic pipeline data, so anyone can use it without sourcing proprietary geodata first
- A real anomaly-detection rule → every segment's expected pressure drop is compared against its actual drop, live, from whatever data is loaded — not hardcoded per scenario
- Investigation trace → clicking a flagged segment highlights the upstream path and the downstream branch it's starving. Issue 05's ‘isolate node ranges, see the blast radius,’ now one click
- Bring your own network → users upload their own nodes/edges CSVs; it runs entirely client-side, so no pipeline data ever leaves the browser
Where the IntelliFlux tool proved the idea inside one company — and outlived it — Manifold is the same thinking made into something anyone can open and try.