Applied AI · Robotics · Engineering

From models
to things that
work in the world.

I'm Praveen, a software engineer with a mechanical engineering foundation. I build AI systems, test their limits, and connect them to real workflows and physical machines.

Explore the work ↓View résumé ↗
12+ years in production softwareMS Mechanical Engineering, ASUMBA University of Washington

Selected projects

What I built. What I learned. The evidence behind it.

01 / Physical AI

SortBot

A self-funded warehouse robot, from CAD and 3D-printed hardware to learning from demonstrations and testing in a working warehouse.

I built a dual-arm SO-101 sorting station, trained an Action Chunking Transformer policy with LeRobot, and integrated the learning pipeline with warehouse systems.

Imitation learningACT / LeRobotTeleoperationCAD
Watch the warehouse demonstration ↗
What the warehouse test taught me

Flexible objects such as plastic bags exposed limits in the learned behavior. The experience led me to questions about perception, gripper design, and feedback: how can a robot adapt when the object or observation changes?

DEMONSTRATE → TRAIN → OBSERVE → ACTConceptual illustration · demo linked alongside
02
Post-training & evaluation

Adapting an open model
for real workflows

At Flexport, I moved from an AI hackathon prototype to taxonomy classification, model evaluation, and inference deployment work.

QLoRA / SFTGRPOPyTorchvLLM
  • Built a reusable training and evaluation pipeline that won Flexport's 2026 Summer Hackathon.
  • Fine-tuned Qwen on operational data and evaluated checkpoints in the full agent.
  • Worked with colleagues and Baseten to compare inference configurations, latency, throughput, and cost.

The lesson: a higher overall accuracy can still hide more costly, high-confidence errors. Evaluation needs to reflect how the system is used.

Read the Flexport Engineering write-up ↗
03
Computer vision

Port Visibility

Exploring how video can provide auditable evidence of container-transfer events at ports, with timestamps and uncertainty.

Open VLMsOpenCVObject tracking
  • Built a timestamped clip-capture pipeline and evaluated camera viewpoints.
  • Experimented with Molmo, InternVL, VideoLLaMA, and other vision-language models.
  • Combined stabilization, fence masking, motion detection, tracking, and crane-geometry rules.

The question: when the footage does not support a conclusion, how should the system express uncertainty or seek a better view?

Explore footage and project artifacts ↗
04
Agents & production engineering

Boomi Code Review Assistant

An AI-powered review platform for integrations built on Boomi, developed and shipped at Flexport.

PythonClaudeXMLAWS
  • Parsed integration XML into architecture graphs for review.
  • Identified bugs, dependency risks, and security deviations.
  • Added an evaluation judge and deployed the platform for the Ecosystem team's production reviews.

200+ production reviews supported, as documented in my résumé.

See engineering experience ↗
05
Scientific computing

Modeling physical systems

Research in turbulence, numerical methods, and feedback control shaped how I approach computational problems.

C / FORTRAN / MATLABMPI / OpenMPControl

Arizona State University: under Professor Ronald Adrian, I applied linear stochastic estimation to turbulent-flow structures, integrated it with direct numerical simulation code, and used parallel computing.

IIT Madras: under Dr. Prasad Patnaik, I developed a numerical solver for the Ginzburg-Landau equation and implemented PID and fuzzy-logic feedback control.

Read my ASU master's thesis ↗

The artifact shelf

Demos, footage, writing, and research in one place.

SortBot warehouse demonstrationVideo collection · Google Drive
Port Visibility projectFootage and supporting artifacts · Google Drive
Fine-tuning for taxonomy classificationTechnical write-up · Flexport Engineering
Dynamics of vortices in numerically simulated turbulent channel flowMaster's thesis · Arizona State University
Full engineering résuméExperience, education, and technical skills
Behind the work

Software experience.
Physical curiosity.

My background spans mechanical engineering, scientific computing, and more than a decade building production software. I enjoy following a problem across layers: from the application and model to deployment, sensing, and hardware.

I'm interested in software engineering and applied AI opportunities, and in the questions that connect perception, learning, and robotics.

FlexportSenior Software Engineer · Applied AI & integrations
Expedia GroupSoftware engineering · Platforms & reliability
FlexeSoftware engineering · Logistics integrations
Wolfram ResearchQuality engineering · Computational software