Skip to content
Role
Product · UX · Frontend (solo)
Team
Solo
Period
2026.05 – 2026.06
Tools
Next.js, TypeScript, Supabase, Leaflet, Tailwind, Figma

The web console that watches the sickness metric

Glide Console

Glide Console running live – district scores, map layers and the event feed

The research tool for Glide, built to test the UX before the algorithm existed.

Summary

With no algorithm yet, the warning and the silence still had to be tested. So the tool came first: a person sets the timing and the intensity of every intervention by hand while the rider sees what looks like an automatic system. Sensors, simulation, audio, haptics and vision are tuned on one screen.

5
Things tuned on one screen

Sensors · simulation · audio · haptics · vision

6
Map layers

Output · heatmap · districts · roads · incidents · signals · crosswalks

27,763
OSM points on the map

Output · technical scale. 6,793 signals · 20,970 crosswalks

Overview

  • Experiment control – a person picks when to intervene, through which channel, and how strongly. The rider's screen looks exactly like an automatic system.
  • District scores – a weighted average for the whole city, 0–100, higher is safer, with all 25 districts sorted into safe, watch, rough and severe.
  • Risk top five – the lowest-scoring districts, in order. This is the screen you route around.
  • Active rides – current rides split across live, simulation, game and idle.
  • Event log – hard braking, sharp turns, recovery actions and route comparisons, timestamped and tied to a participant code so a session can be replayed later.

Problem

  • Waiting for the algorithm was not an option – warning times, the threshold for silence and the switch between senses all depend on what a person feels. Those answers were needed before any code could make the call. Having a person reproduce the intervention by hand (Wizard of Oz) tested the UX ahead of the algorithm.
  • Why one screen – validating the sickness model meant sensors, simulation, audio, haptics and vision (camera) being tuned in step. Across five separate tools, one ride cannot be reproduced under the same conditions twice.
  • Now – it started as a research tool. After the pivot, the console became a product asset for B2B customers.

Key decisions

  1. 01Build the tool before the algorithm

    Finishing the code that decides automatically and only then learning that the UX was wrong is the expensive order to work in.

    →An experiment console where a person operates the timing and the intensity came first. It carries session control, an MISC timeline, live sensor logging and a sickness profile per participant. To the rider it looks automatic.

  2. 02Put all five on one screen

    With sensor readings, simulation, sound, vibration and camera in separate windows, one ride cannot be rebuilt under the same conditions.

    →Ten menus live in one console: control, routing, sensors, virtual drive, the cue system, IVI, live drive, game drive, operations and analysis (output). The information density was not lowered to consumer levels. One operator reads this screen.

  3. 03Canvas instead of SVG, so everything can be on at once

    Drawing 6,793 signals and 20,970 crosswalks as DOM nodes freezes the browser.

    →One canvas renderer draws all 27,763 points in a single pass (output). District polygons stay on their own SVG layer above it, so the pointer always hits a district boundary first.

  4. 04Dark map, light panels

    The map carries its data in colour, and darkening the panels too makes numbers and colour compete.

    →In dark mode only the map darkens; cards and the sidebar stay light.

The sensor, audio and vision modules

Sensors – with no hardware attached, mock IMU readings take over automatically. The same screen runs without the board.

Simulation – a drive runs on a canvas, turns are detected from the change in bearing, and warning cues are generated from them. The same stretch can run with Glide on and off for comparison.

Audio – a 100 Hz sync signal, boosted past a lateral-acceleration threshold, with the Arduino motors firing at the same time.

Vision – a webcam reads head posture and hand-to-mouth gestures and escalates the recovery step by one. The video is never stored.

Process

The map

Six layers stack on the map: a sickness heatmap, district polygons, live road speeds, incidents, traffic signals and crosswalks, fed by TOPIS traffic, Tmap roads and OSM/Overpass.

Data

One Supabase project holds seven tables: participants, sessions, events, the sickness time series, haptic firings, vehicle positions and district scores.

Every session carries a mode – live, simulation, game or idle – so real drives and simulated ones replay through the same screen. Without the sensor board, faked readings drive the same view.

Results

  • The flagship numbers came from here – this console's simulation engine (simDriveEngine.ts) ran the same stretch with Glide on and off and compared the accumulated MSDV curves. The sickness-metric figure in Glide's results (−77% · simulation) is that comparison. No in-car controlled trial was run.
  • Four axes on one screen – physical (IMU-based MSDV), physiological (heart rate, head posture), subjective (MISC, FMS) and behavioural (gestures read by camera). An effect only counts when all four agree. All four are built, but the sample is too small for statistical significance.
  • Deployed – running on Vercel production. The film on this page was shot after the three fixes below.

Reflection

What held up

  • Building the tool before the algorithm. Whether the warning and the silence worked was answered before any code decided it. And the operations screen built for experiments became the product asset a customer would use.
  • Making the screen run without the board. When the hardware dropped out, neither the study nor the demo stopped.
  • Keeping the schema in the repo. After the database disappeared, it went straight back up on a new project.

What failed

  • I thought shipping was the end. Four months later the screen was empty. Three things had gone wrong at once.
    • The database was gone. A free-tier project that had sat idle was deleted – not unreachable, the hostname itself no longer resolved.
    • The map provider changed its terms. The keyless tiles started arriving stamped "API KEY REQUIRED". Swapped for a provider that needs no key.
    • One condition was inverted. The district panel had always read zero, not because of the data but because of if (!cancelled || !data) return; – it bailed out even on a good response.
  • I suspected the data before the code. The cause was one line.

What I would test next

  • A larger sample across the four axes. The axes exist; the participants do not.
  • A direct link between the board and the app. It goes through the web today; direct Bluetooth is untouched.
  • Where else this still leans on somebody else's free tier.