arrow_backBack to What I Build

Computer Vision

Your warehouse camera sees everything. Your software sees nothing.

Most camera AI only works in perfect lighting. Real businesses deal with shadows, blocked views, moving water, and objects that do not sit still. I build systems that work where the work happens — extracting 491 measurements from a person, counting tiny oyster seed, or inspecting gear fast enough to keep up with the team.

What Usually Looks Normal Until I Point At It

A person counting or inspecting things by hand, all day, with a known margin of error nobody has fixed

Off-the-shelf vision tools that work in the demo and fail in your warehouse

Latency too high for real-time use: models that run in seconds, not milliseconds

Privacy concerns send sensitive images to third-party APIs

How I Build It

01

Start with the specific detection, measurement, or identity problem

02

Select the vision approach by accuracy, latency, and the kind of scene the business actually has

03

Build pipelines that handle real-world variation and occlusion

04

Deploy on owned hardware or edge devices based on latency and privacy needs

Outcomes

491-field face/body inventory for style, wellness, or identity applications

Instant subject isolation in video

Technology Stack

PythonBody-inventory backbonePrompt-driven segmentation308-keypoint poseFast 3D bodyOpen-vocab detectionOpenCVPyTorch

Capabilities

Face & Body Analysis

A single system reads pose, outline, and depth at once. It measures the face and creates a unique body profile—checking face shape, symmetry, body type, posture, stance, and specific measurements.

How It Works

  • Unified backbone extracts pose, segmentation, depth, normal, pointmap
  • Face analyzers measure shape, symmetry, thirds, subregions

Outcomes

  • 491-field face/body inventory per subject
  • Body-twin matching via 45-dimensional shape fingerprint

Detection, Segmentation & Tracking

I combine segmentation, detection, measurement, and tracking — calibrated for your camera, lighting, and throughput.

How It Works

  • Select the detection path by task: masks, counts, measurements, or tracking
  • Calibrate pixel-to-world measurements for the deployment camera

Outcomes

  • Real-time counts in variable field conditions
  • Pixel-accurate segmentation for measurement and masking

Related Case Studies

The research is public.

Benchmarks run on machines I own, measured against frontier systems — with the failed approaches reported next to the results. Read it before you take my word for anything.