Dhruva Shekdar

Building AI systems for robotics and autonomous vehicles.

I build AI systems for real-world constraints: offline execution, efficient model adaptation, operator-facing workflows, and robotic perception.

View resumeSelected work

500+

production scans per day

Tesla Android workflow used by operators for cell-material verification at Giga Texas.

95%+

fewer trainable parameters

Reported for JoLA fine-tuning of Llama-3.2-1B in Critical ML Lab experiments.

89,768

high-resolution frames

Prepared for the Mini-DLSS temporal video super-resolution training pipeline.

01 / Experience

Built under real constraints.

Production AI, language-model research, factory-floor software, and robotic perception.

2026

Ford Motor Company

AI / Machine Learning Intern, A.I. Team

Developed resilient in-vehicle AI infrastructure spanning offline inference, audio ingestion, and structured vehicle-tool invocation.

  • Architected a six-phase fallback system using Python and ONNX Runtime.
  • Designed dual-stream audio ingestion for online and embedded models.
  • Implemented typed MCP function calling with schema validation and permission gating.

Python · ONNX Runtime · MCP · On-device AI

2026 - present

Critical Machine Learning Lab

Undergraduate Machine Learning Researcher

Researching parameter-efficient model adaptation, model evaluation, and reproducible experimentation for language models.

  • Fine-tuned Llama-3.2-1B with JoLA, with reported loss moving from 2.03 to 0.69.
  • Evaluated experiments across five NLP benchmarks.
  • Built a sweep runner for more than 100 configurations and contributed ReFT integration work to TamperBench.

PyTorch · LLM evaluation · JoLA · W&B

May - Sep 2025

Tesla

Software Developer Intern, Cell Material Flow

Built offline-first Android software for high-volume material verification workflows at Giga Texas.

  • Supported more than 500 structured scans per day across 12 operators.
  • Reduced material verification time by 75%.
  • Analyzed more than 200 tracking failures; daily mismatches fell from 13 to 2.

Kotlin · Java · Android · gRPC · MVVM

Dhruva Shekdar during his Tesla internship at Giga Texas

2026 - present

WATonomous

Machine Learning Engineer, Humanoid Robotics

Developing visual perception for autonomous keyboard interaction in a ROS-based humanoid robotics stack.

  • Training and evaluating YOLO-based keyboard detection models.
  • Using RGB-D input for keyboard detection and 3D pose estimation.
  • Integrating perception outputs into the robot autonomy stack.

PyTorch · CUDA · RGB-D · ROS · YOLO

02 / Selected work

Systems that perceive, adapt, and operate.

A focused set of projects selected for technical depth, measurable evaluation, and real-world integration.

Case study 01

LLM evaluation and grounded clinical QA

FDA-Gemini Benchmark

An evidence-first benchmark for testing whether LLM agents can reason over FDA drug labels without losing citations, qualifiers, or refusal discipline.

Profiled the source corpus, built 12 Harbor task suites with hidden deterministic verifiers, introduced fractional scoring, and analyzed model failures across retrieval, dosage, synthesis, comparison, and refusal tasks.

FDA labels profiled
704
Harbor task suites
12
scored hard-set trials
23

Python · Harbor · LLM evaluation · Gemini

Read case studyView repository
Bar chart showing mean fractional reward across seven FDA benchmark task families
Bar chart showing content and clinical qualifiers as the dominant benchmark failure criteria
Independent research benchmark. Results describe recorded model runs and are not clinical guidance.

Case study 02

Temporal video super-resolution

Mini-DLSS

A compact BasicVSR-style pipeline for reconstructing high-resolution video from multi-frame context.

Built the data preparation, temporal model training, quality evaluation, and CPU latency benchmarking pipeline to study quality, temporal stability, and model-size tradeoffs.

PSNR-Y
38.33 dB
over bicubic
+1.52 dB
ONNX CPU / frame
21.59 ms

PyTorch · ONNX · OpenCV · Vimeo-90K

Read case studyView repository
Mini-DLSS comparison strip showing low-resolution input, reconstruction, and reference frames
Mini-DLSS quality and latency comparison across bicubic, single-frame, and temporal methods
Results are measured on a local Vimeo-derived REDS-style validation set, not the official REDS benchmark.

Case study 03

Autonomous vehicle perception and control

Hack the Move

A first-place hackathon system that connected visual perception to an autonomous vehicle control stack.

Trained stop-sign and traffic-light perception models, then converted detections into desired-speed overrides consumed by an existing Simulink controller.

overall placement
1st
reported mAP@50
91.3%
stop command target
0 m/s

Python · YOLO · MATLAB · Simulink

View repositoryWatch demo
Engineering team working on the autonomous vehicle during Hack the Move
Keep precision, recall, mAP, and accuracy distinct in any expanded case study.

03 / About

Software that meets the physical world.

I study Mechatronics Engineering at the University of Waterloo with an Artificial Intelligence option. My work sits where model behavior meets system constraints: latency, offline operation, noisy sensor input, hardware integration, and measurable user outcomes.

I am most useful on teams building AI products that must work outside a notebook, especially robotics, autonomy, on-device inference, computer vision, and ML infrastructure.

Languages
Python · C++ · Java · Kotlin · Bash
ML and vision
PyTorch · ONNX Runtime · OpenCV · CUDA
Systems
ROS · gRPC · Android · Linux · Docker

04 / Contact

Building an AI system that has to work in the real world?

I am interested in ambitious engineering teams working across AI, robotics, autonomy, and high-performance software.

Start a conversationReview resume