Albert Jordan
Porduct Management, Ecosystem Development, Team Building
Porduct Management, Ecosystem Development, Team Building
Dynamic and results-driven software products executive with deep expertise in mobility, automotive, and AI-powered solutions. Skilled at building and leading high-performing teams across both large enterprises and fast-paced startups. Proven ability to deliver innovative, market-leading products by blending strategic vision with hands-on technical execution. Adept at navigating ambiguity, streamlining internal processes, shaping product architecture, and cultivating strong partner ecosystems.
Calm and decisive under pressure, with a “roll-up-your-sleeves” leadership style that ensures products move efficiently from concept to launch. Experienced in negotiating complex technology contracts and driving sustainable business growth through innovation and collaboration.
Ready to tackle new challenges and continue to bring compelling automotive experiences to market.
Personalized Real Estate Agent
This is the final project for Udacity’s Gen AI nano degree course.
The project implements a real estate listing application. In an industry where personalization is key to customer satisfaction, the ficticious company wants to revolutionize how clients interact with real estate listings. The goal is to create a personalized experience for each buyer, making the property search process more engaging and tailored to individual preferences.
The task is to develop an innovative application named “HomeMatch”. This application leverages large language models (LLMs) and vector databases to transform standard real estate listings into personalized narratives that resonate with potential buyers’ unique preferences and needs.
There are two components, implemented in Google Colab notebooks. Colab was selected because the solution required GPU for image generation.
In process_Listings.ipynb I first use openAI to get a list of neighborhoods in San Francisco. Then for each neigborhood I come up with a ficticious listing, and a description that is based on the characteristic of the selected neighborhood. I also use the description to generate an image of the property. Below are two samples:
The second component if the find_home.ipynb. Given a list of customer preferences, the code searches for properties that match the user request. The description of the property is modified to match the user preferences. Also, the images are vectorized, so the code searches the images for the description. For example, if the user had indicated that they wanted a property with good views, the code searches the images for that and comes up with the following options:
AI Assisted Photo Editing
In this class we have seen how Generative models can be used in Computer Vision. We have also seen how to use the Segment Anything Model to select subjects in images by providing points and other inputs.
Let’s put some of this knowledge to good use by building something interesting, and fun!
This project implements a little app that allows you to select a subject and then change its background, OR keep the background and change the subject.
The process involves a user uploading an image and selecting the main object by clicking on it. The Segment Anything Model (SAM) is activated to create a mask around the selected object, choosing the most accurate mask generated. The user is shown this result to either accept it or refine the mask further with additional points.
Once the mask is finalized, the user gives a text description (and possibly a negative prompt) to specify a new background for the selected object. An infill model then creates this new background, and the final image is displayed. Optionally, the user can choose to invert the mask and substitute the subject while keeping the background, as in the example above.
The app can be used to swap backgrounds, swap subjects, remove objects, etc.
There is a file in github repository that has some samples of the output of the app. Below is a sample. Note that two options, where the selected item is kept, and a new background is generated using Gen AI, or conversely, the seleted item is taken away.
LLM based Modern User Manual
Having learned a few things about chatbots, I decided to do a fun project. For me, one problem I have when driving a new car is seeing telltale signs in the dashboard, and not knowing what they are. I wanted to do a chatbot that would address this pain point but also wanted to incorporate images.
Picture below depicts the data processing piece of the project. I scraped the vehicle user manual (I did both for Tesla and Lucid Air). For each page that was scanned, I looked for both images and text. My aim was to generate a vector, for each paragraph enountered. I would also couple paragraphs before and after an image, and feed the text and the image to openai to come up with a more context aware image caption. I would generate an embedding for that, and would store both types of vector in a pinecone database.
The Chatbot piece was not difficult, except that I had to account for processing both images and text, and ensure that the images followed relevent text.
Convinced NVIDIA to design a Tegra 3 compute fabric. Led prototype design and implementation of a 1-DIN head unit prototype, with replaceable display and compute elements and installed in 20 cars. Negotiated a $100K evaluation agreement with LG, which has led to a commercial development opportunity.
Maintenance Tool
Although my primary responsibility at Nikola was to drive the strategy and program for ADAS, the project that was more consequential as development of a chatbot to help troubleshoot trucks. As I attended calls and meetings, what struck me was that there are a number of Jira projects created by different groups (e.g. development, quality, manufacturing,… seven in total) and number of other databases that housed information on existing experience with the fleet. However, these projects were very poorly maintaned.
The tendency of the team was to create a new Jira project, port some tickets from existing projects to help with truck diagnosis. My hypothesis was that we needed to salvage any information that was captured, and if the tickets were poorly maintained, then one could use LLMs to fill in the holes. As an example, Jira tickets had fields for problem, root cause, interim solution, long term solution, but these fields were routinely left blank. Instead, the comments on the tickets would hold infomation on such items.
I used openAI to scrape all the jira projects, as well as other database, to produce a table that essentially held “problem”, “Cause”, and “Solution”, for each logged ticket. I then combined “problem”, and “cause” fields and produced a vector that was stored on Pinecone, a vector database.
Once the Pinecone database was created, I used the pipeline structure below to develop a pipeline for a RAG based chatbot.
Although initially, I had used Jupyter notebook to implement the chatbot, for better affect I developed a very simply Flask application that allowed me to surface a UI that would be similar to what a user would expect from a chatbot.
Below you can see a sample output. Given a question, the chatbot would come up with potential causes, reference existing tickets and vehicles where the problem was observed. I got sample questions from the maitnenance team, and the results were suprisingly good.
The tool was considered a priority, and I was asked to implement the chatbot given Nikola’s current AWS framework. I ported my database to a Postgres database, and instead of OpenAI I used anthropic to implement the chatbot.
I also used the vectorized database to do an analysis of current problems faced by Nikola in an attempt to develop a prioritized list of target areas.
layout: default categories: project modal-id: 14 date: 2024-11-01 img: hugging.png alt: image-alt project-date: December 2024 client: N/A category: Foundation Models, Gen AI description: Parameter Efficient Fine Tuning —
Industry’s first commercial Android Auto receiver. Negotiated a $500K POC contract with Hyundai, Honda. Managed delivery of a projected mode POC showcased at CES. Led the discussion with Google and Hyundai to deliver an OAA compliant receiver architecture that leveraged the POC. Managed team of engineers to deliber the receiver that led to the commercial deployment of the fist Andrid Auto solution in the industry.
Managed the engineering team that Delivered a projected automotive experience at Jaguar Landrover, an early example of voice first experience. Developed requirements that allowed transition of a demo to a commercial grade service. Drove development of standalone validation tests to expedite integration with Nuance, Bosch, and HERE SDKs. The project was a big success, with JLR customers indicating the projected experience was superior to the embedded experience. This led to creation of a project that led to a $15M investment in CloudCar, and a partnership with LG, to deliver the experience.
Convinced NVIDIA to design a Tegra 3 compute fabric. Led prototype design and implementation of a 1-DIN head unit prototype, with replaceable display and compute elements and installed in 20 cars. Negotiated a $100K evaluation agreement with LG, which has led to a commercial development opportunity.
Parameter Efficient Fine Tuning
This project was part of Udacity’s Gen AI Engineering Nano Degree program. Lightweight fine-tuning is one of the most important techniques for adapting foundation models, because it allows you to modify foundation models for your needs without needing substantial computational resources.
In this project, brings together all of the essential components of a PyTorch + Hugging Face training and inference process.
The project:
Loads a pre-trained model and evaluate its performance Perform parameter-efficient fine tuning using the pre-trained model Perform inference using the fine-tuned model and compare its performance to the original model
Udacity Nano Degree
### In progress
Udacity Nanodegree
Description: The Natural Language Processing (NLP) Nanodegree program is designed to equip students with the skills necessary to enable computers to understand, process, and manipulate human language. This advanced program focuses on practical applications, allowing students to build models using real data and gain hands-on experience in various NLP tasks such as sentiment analysis, machine translation, and more.
The program includes several projects that allow students to apply their learning in practical scenarios. While the specific projects may vary, they typically cover the following areas:
Udacity Self Driving Car Engineering Nanodegree
The Self Driving Car Engineer Nanodegree program at Udacity is designed to provide you with the skills and knowledge necessary to work on autonomous vehicles. You will learn about the various components that make up self-driving technology, including computer vision, deep learning, sensor fusion, localization, and control systems. The program combines theoretical knowledge with practical projects, allowing you to apply what you’ve learned in real-world scenarios.
The program typically includes the following projects:
CarND Term1 Starter Kit: Set up your development environment and get familiar with the tools and libraries used in the program.
Traffic Sign Recognition: Build a model to detect and classify traffic signs using computer vision techniques.
Lane Detection: Implement a lane detection algorithm that identifies lane boundaries in images from the car’s camera.
Behavioral Cloning: Train a deep learning model to drive a car by mimicking human driving behavior based on video input.
Advanced Lane Finding: Enhance your lane detection project to handle more complex scenarios, such as curved lanes and variable lighting conditions.
Object Detection: Use deep learning techniques to detect and classify objects in images, such as pedestrians and other vehicles.
Sensor Fusion: Combine data from multiple sensors (e.g., cameras, Lidar, radar) to improve the accuracy of the vehicle’s perception.
Path Planning: Implement algorithms to plan a safe and efficient path for the vehicle to follow in a simulated environment.
Final Project: Integrate all the skills and knowledge you’ve gained throughout the program to create a comprehensive self-driving car system that can navigate a complex environment.
Computer Vision Engineering Nanodegree
Computer Vision Nanodegree program at Udacity is designed to help students master the skills needed for advanced applications in robotics and automation. Here’s a brief overview:
Duration: Approximately 2 months Difficulty: Advanced Summary: This program focuses on teaching the essential computer vision skills that are critical for modern technology applications. Students will learn to analyze images, implement feature extraction, and recognize objects using deep learning models. The curriculum includes hands-on projects that allow students to apply their knowledge in practical scenarios. Skills Taught The program covers a variety of skills, including:
Image Classification: Building a model to classify images into different categories.
Object Detection: Implementing algorithms to detect and localize objects within images.
Facial Recognition: Developing a system to recognize and verify faces in images.
Managed the engineering team that Delivered a projected automotive experience at Jaguar Landrover, an early example of voice first experience. Developed requirements that allowed transition of a demo to a commercial grade service. Drove development of standalone validation tests to expedite integration with Nuance, Bosch, and HERE SDKs. The project was a big success, with JLR customers indicating the projected experience was superior to the embedded experience. This led to creation of a project that led to a $15M investment in CloudCar, and a partnership with LG, to deliver the experience.
iOS App Development
Since 2010, I have regularly followed the CS193p class at Stanford. Understanding iOS development helps me keep at the forefront of state of the art software development paradgims. I started with learning about Objective C program and Model-View-Controller software development process, to learning about Swift and SwiftUI and functional program. Some of the projects are below.
https://github.com/albertjordan/Calculator
Most recent CS193 courses have begun to dive into SwiftUI and funcitonal programming. Here is the latest class project with a memory game, where the user is given a deck of cards and the objective is to match two cards with the same image. https://github.com/albertjordan/Memorize2023
Some of the projects I have done that leverage Objective C devepment include automotated test framework for a visual voicemail service deployed at Sprint, a video communication automated video quality test bend.
For the visual voicemail, I created a automated test framewoked that leveraged Skype and an amazon voice to text service. The framework would generate call using skype, where by date stamps and a sample message would be send to a number. Connecting to an Android phone thru the adb interface, the framework would ensure voicemail were received in the right order, correctly and accurately.
For the video quality test framework, I developed a test framewok whereby two Android phones were connected thru adb. One of the phones would initiate a video call to the other phones. Both phones were connected via Wi-Fi access point on a mac. Then the data rate within the Mac Router would be managed to inject errors, or throttle the bandwidth, and framework would monitor the Frame Error Rate at the receiving Android phone via ADB to ensure they met an accessptable user experience criteria (frame rate with acceptable error rate).
CloudCar CTO had proposed a novel approach to architecting a head unit. The thought was that the main pain point for OEMs was ensuring the in car experience was on par with consumer’s phone experience. But whereas consumers switch phones every 2 or 3 years, the lifespan of a vehicle is much longer.
CloudCar’s proposal was to divide the head unit into two components: a baseband, and a compute module. The baseband included elements that were constant with time. For example, tuners, amplifiers were items that were basically constant with time. On the other hand, the compute fabric needed to kep pace with mohr’s law.
I convinced NVIDIA to design a Tegra 3 compute fabric for cloud car. I led prototype design and implementation of a 1-DIN head unit prototype, with replaceable display and compute elements
I used an off the shelf 1-DIN am/fm radio as the base band, and came up with a 2-DIN solution that could be insalled in vehicles. I installed the solution in 20 cars. Below, the picture of a modular head unit in my Sienna van.
Explored feasibility of leveraging large language models, and RAG based technology for Grand River Solutions (GRS). GRS Helps create safe and equitable work and educational environments and brings systemic change to how school districts and institutions of higher education address their Clery Act & Title IX obligations.
Developed prototype to prove feasibility of LLMs and chatbot in automating generation of formal interview reports. Demonstrated time and cost savings on sample use cases. Authored detailed requirements for an off-shore developement team to implement a No/Code solution based on Bedrock for initial application.
Industry’s first commercial Android Auto receiver. Negotiated a $500K POC contract with Hyundai, Honda. Managed delivery of a projected mode POC showcased at CES. Led the discussion with Google and Hyundai to deliver an OAA compliant receiver architecture that leveraged the POC. Managed team of engineers to deliber the receiver that led to the commercial deployment of the fist Andrid Auto solution in the industry.
Prior to iPhone launch in 2007, the wireless industry struggled with delivering compelling applications to consumers. I worked at Core Mobility, which was one of handful companies that was able to deploy applications on consumer grade wireless feature phones.
I started with leading the product management team, but in time my responsibilities grew to include the management of the handset engineering team, QA, and in a few occasion as leading engineering. Core Mobilities competency was development of a process that allowed bug-free integration with wireless phone platforms.
The services I delivered were wirlesss backup (which allowed seamless phonebook restoration for lost or new phones), voice messaging, and visual voicemail (experience that is available on iPhone today.)
Responsible for lucent’s CDMA infrastrcuture product management team, a $2B/Year business at the time.
Led a team of 40 product managers, with product covering cellular and pcs customer sets.