Joanna
Hoang

The situation

CBP wanted to identify everyone in a moving vehicle at secured entry points. NEC's NEVA system triggers several high-resolution cameras as a car passes and sends the best face image of each occupant to a biometric matcher.

My role

Hands-on for about nine months of the pilot: test-data collection and cleaning, configuring test environments, validating the biometric API and matching workflow, and running repeated field tests.

The approach

The controlled demo worked; the field didn't. Night lighting, rain, tinted windows, camera angles and moving vehicles all hurt capture. In one test round I coordinated a same-day fix with portable event lighting and pop-up canopies over the capture zone. Matching recovered, which proved the environment was the cause, and engineering designed the permanent solution.

The outcome

That evaluation is what the customer bought on. The deal landed and the pilot rollout began; full rollout continued after I left.

• ~9 mo hands-on in the pilot
• 1 day to prove the field problem was the environment

The situation

A federal customer came to NEC asking about facial recognition. Discovery showed the real need was long-term intrusion detection around a large, rugged outdoor perimeter, with little capacity for maintenance.

My role

Gathered requirements with the customer and operators and worked through fit with engineering, who made the final recommendation. I explained it to the customer, coordinated through installation, built the operator dashboard and stayed on for post-sales support.

The approach

IPIDS turns buried or fence-mounted fiber into the sensor, with AI that tells climbing, cutting and digging apart from rain, wind and animals, and no powered equipment along the fence. I built the dashboard on a real-time event platform, similar in role to AWS IoT: each intrusion as it happened, what was detected and when, the live camera feed for the alert area, and charts of activity over time. When events stopped arriving after an environment migration right before a demo, I traced the path and fixed the network configuration with the platform's engineer.

The outcome

The system was deployed and I supported it after the sale. IPIDS itself won the 2023 ASTORS Platinum award for Best Perimeter Protection Solution.

The situation

NEC's Smart Access visitor-management system needed to send candidates to a third-party background-vetting provider and get a pass, fail or needs-review decision back.

My role

Wrote the integration contract in OpenAPI 3.0, covering the provider's side from its integration guide and Smart Access's side from our system diagrams, and defined the vetting scenarios.

The approach

An asynchronous job pattern: submit, poll status, retrieve the result, with an optional callback. Twelve scenarios covered 2- and 3-point identity checks, with or without a photo, against criminal-history and DMV records plus a face match. My main design call: key each vetting request to a transaction ID instead of the person, because one candidate can be vetted more than once.

The outcome

A validated OpenAPI 3.0 contract for both systems, with Java models generated from it using OpenAPI Generator for the build.

The situation

Enterprise teams had data-science skills but little practice with distributed training or Intel's optimized AI frameworks, which slowed adoption of Intel's AI hardware.

My role

Ran the AWS EC2 side of the hands-on PyTorch and TensorFlow optimization labs and prepared their instructions, scripts, notebooks and Docker images. Staffed the Deep Learning Boost demo booth and supported the AI Reference Kit hackathon.

The approach

Labs at Intel Innovation 2022 and developer-summit webinars; a two-node PyTorch distributed-training lab on AWS that let developers watch training spread across machines; and the AI for Social Good hackathon, where 100+ developers built on Intel AI Reference Kits. Each environment had to hold up with hundreds of people using it at once.

The outcome

Over 30 demos, workshops, conference labs and proof-of-concept sessions, reaching hundreds of lab participants and more than 100 hackathon developers.

• 30+ sessions
• 100+ hackathon developers
• 100s of lab participants

The situation

Intel packaged AI workflows into catalogs developers could run on their own, such as ResNet-50 transfer learning for image classification on sports, medical histology and satellite imagery. Anything that broke on a developer's first try would cost adoption.

My role

The customer's stand-in, working with Intel's AI development teams. I didn't write the original workflows; I made sure they worked for the people who would use them.

The approach

Ran the vision transfer-learning workflow end to end the way a developer would, reported every breakage, validated the Docker image and conda environment, ran the sample and benchmark runs, and updated the documentation to match what actually worked.

The outcome

The workflow shipped in Intel's public AI workflows catalog. Separately, I validated Intel's MLPerf benchmark results on AWS EC2. See the published workflow ↗