T-Minus 4 Weeks: Rapid Prototype for Spatial Object Tracking Product

The Situation:

An AI startup needed to secure a large commercial contract with a Fortune 100 healthcare company. To do so, they had four weeks to demonstrate accurate, near–real-time tracking of people moving through a physical environment—measured in milliseconds, not seconds.

This was not a science project. The demo needed to work reliably, under executive scrutiny, with no room for hand-waving.

We had three things working in our favor:

• a clearly defined success criterion

• a non-negotiable timeline

• a small, focused team with explicit ownership

That clarity shaped every technical decision that followed.

The Approach:

Rather than over-optimizing models or chasing edge cases, we designed the prototype around a simple question:

What is the minimum system that convincingly proves this can work in production?

At a high level, the system needed to:

1. Capture positional data from physical sensors

2. Identify people in the environment

3. Translate noisy sensor data into stable trajectories

4. Visualize movement in near real time

Each step was chosen to de-risk the next.

Key Technical Decisions:

• Physical environment & instrumentation We converted an office space into a controlled mock retail environment and instrumented it with video sensors. This allowed us to test realistic movement patterns while keeping variables manageable.

• Spatial alignment We anchored sensor output to a shared coordinate system using physical reference markers, projecting all detections onto a unified 2-D plane. This made downstream reasoning and visualization tractable from day one.

• Person detection We integrated a person-detection model directly into the sensor pipeline, mapping detections to approximate real-world positions in near real time.

• Noise handling (where most prototypes fail) Real-world sensor data is messy. Raw output caused erratic trajectories—people appeared to jitter and teleport across the space. Instead of smoothing after the fact, we applied a Kalman filter to model both motion and noise explicitly. This preserved responsiveness while producing stable, believable movement—critical for executive trust in the demo.

• Visualization We streamed the processed data into a live visualization, making latency, movement speed, and positional accuracy immediately legible. A parallel front-end effort delivered a custom UI that overlaid movement on a floor plan and tracked session-level metrics.

Outcome:

This was a prototype, not a full product. Some production features—like robust re-identification—were intentionally deferred.

What mattered was that, within four weeks, the system could:

• Detect people moving through a physical space

• Estimate and smooth their trajectories

• Stream accurate positional data to a live dashboard in near real time

The demo landed. The executive team approved the engagement, resulting in the largest commercial contract in the company’s history. The prototype became the foundation for a production system used across multiple clients.

Why this matters:

Many AI demos fail not because the models are weak, but because the system design collapses under real-world constraints. This project succeeded because we treated the prototype like the first slice of a production system, not a throwaway proof-of-concept.

That mindset—designing for reality even under extreme time pressure—is what I bring to every engagement.

________________________________________________

If you’re building an AI prototype that needs to survive contact with real users, real data, or real deadlines, I’m happy to talk.

Start a conversation