Skip to content

// about

Clive Lewis

A coconut-harvesting robot doesn’t sound like the beginning of a software career. But that was my undergraduate capstone as a mechanical engineering student in Mumbai. I wrote the Arduino C that made it move, and somewhere in that code I found the part of the project I couldn’t put down. It wasn’t the mechanics. It was the code. That was enough to pull me out of mechanical engineering and into software for good.

My first real stop was enterprise work. As a Software Engineering Team Lead at Cognizant, I maintained a portfolio of Java integration interfaces for clients including British Airways, working under SWIFT banking compliance. The work was exacting, and it taught me a habit I still rely on. When a failure keeps recurring, the first fix you reach for is usually just a patch on the symptom, not the cause. Finding that cause means stepping back far enough to actually see it.

It’s the same instinct that brought me to New York for a Master’s in Computer Science at Pace University, and it’s why the classes that stayed with me longest were two Computer Vision electives nobody made me take.

I learn by building. With Flutter, with agentic AI architecture, with the internals of a U-Net, I got further by putting my hands on the problem than by reading about it first. That’s also how I use AI coding tools day to day: not to skip the thinking, but to move faster through implementation so my attention stays where it matters, on architecture, judgment, and whether the thing actually works. And through all of it, I try to stay precise about what I’ve built versus what I haven’t. A working prototype is not a production system, and saying so plainly is part of doing the work honestly.

// experience

Where I’ve worked

  1. June 2025 – Present

    Software Development Intern

    Green2Gold Environmental Education Group · Remote

    I build and maintain a Python REST API backend paired with a Flutter mobile client, supporting content delivery and subscription workflows for the organization’s platform, on a PostgreSQL data layer with clean, reusable endpoints.

    I had never written mobile code before this role. I learned Flutter on the job and shipped production features within weeks, working largely independently and iterating on both backend architecture and mobile UI as requirements moved.

  2. January 2025 – May 2025

    Full Stack Web Development Associate

    Keymate.AI Inc. · New York, NY

    A 4-person engineering team reporting directly to the CTO. I owned a Chrome extension feature end to end — design, implementation, backend microservice integration, and production deployment — serving 1,500+ paying customers.

    I integrated the Google Gemini LLM API into the production application for real-time content summarization and owned the full loop: deploying the model, evaluating output quality against real usage, and iterating on direct paid-customer feedback. Alongside that I built CI/CD pipelines with Jest and Playwright so we could deploy continuously without holding our breath.

    Owning a feature for paying customers makes quality concrete. Users notice within hours when something breaks, and that changed how seriously I take testing and monitoring.

  3. January 2022 – April 2023

    Software Engineering Team Lead

    Cognizant Technology Services Ltd. · Mumbai, Maharashtra, India

    I built and maintained 190+ Java-based integration interfaces for enterprise clients including British Airways, holding 100% SWIFT banking compliance across the portfolio — enterprise financial data integration with real regulatory stakes.

    A recurring production issue kept failing cash remittance transactions under SWIFT cross-validation. My first instinct was to patch each failure as it surfaced, and that didn’t hold — they kept coming back. So I stepped back and traced it through the logs to the actual cause: mandatory system ID fields weren’t populating during cross-validation. I designed an automated Java validation routine that fixed it at the source, and that entire class of failure stopped. The lesson stuck: understand why something is failing before you fix what is failing.

    I also worked directly with business and IT stakeholders to integrate a chatbot and remittance systems with SAP S/4 HANA, improving data recovery by 15% — which meant translating what stakeholders actually cared about (timelines, risk, operational impact) into technical decisions rather than implementing a spec.

// research

Research

Published

Knee Osteoarthritis Detection Using Bone Distances

Clive Lewis, Tarun Ramapuram, Juan Shan. IEEE/ACM Conference on Connected Health: Applications, Systems and Engineering Technologies (CHASE 2025). New York City, New York — June 2025.

In plain terms: we measured the geometric distances between bones in knee MRI scans and trained a model to predict how severe a patient’s osteoarthritis is from those measurements. A MATLAB–PyTorch supervised deep learning pipeline over 160+ scans, reaching 76% accuracy and an AUC of 0.78.

This one started as curiosity. I took Dr. Juan Shan’s Computer Vision and Advanced Computer Vision electives because the subject interested me, not because they were required — the arc from classical image processing through to deep learning stayed with me. That interest turned into mentored research with Dr. Shan, and the research turned into a peer-reviewed publication.

To be precise about what this is: genuine applied deep learning on medical images driven by geometric measurement — model design, training, and evaluation. It is not object detection, tracking, or segmentation work, and I don’t describe it that way.

Unpublished · ongoing

SAM-Net

A two-stage approach to image segmentation. A U-Net produces an initial segmentation; a separate ML model then selects the most efficient points from it; those points go into SAM (Segment Anything Model) alongside the original image to produce a refined result.

I got properly stuck on this one. I could see that U-Net worked, but I couldn’t say why — what was actually happening inside the encoder-decoder. So I built visualizations of the intermediate flattened vector representations at each stage of the network, using AI tools to move quickly through implementation so my attention stayed on the concept. Once I could watch feature maps compress and expand through the network, the architecture stopped being an abstraction.

My test for whether I actually understand something complex isn’t whether I can describe it. It’s whether I can predict what an intermediate step will show before I look.

// education

Education

  • May 2025

    MS, Computer Science

    Pace University, Seidenberg School of Computer Science and Information Systems — New York, NY

    GPA 3.86 / 4.0

    Artificial Intelligence, Algorithms & Data Structures, Parallel Computing and Distributed Systems, plus Computer Vision and Advanced Computer Vision as electives.

  • June 2021

    BE, Mechanical Engineering

    Mumbai University, Don Bosco Institute of Technology — Mumbai, India

    Capstone: a coconut-harvesting robot, programmed in Arduino C. It worked, it was my first real taste of writing code that moves something in the physical world, and I have not claimed to be an embedded engineer since.

Get in touch

Open to Full-Stack, ML/AI, and Mobile engineering roles in the NYC metro area. The fastest way to reach me is [email protected].