My Approach

A large e-commerce retailer's order data was consumed by multiple business domains. A legacy pipeline existed but had unclear ownership, and the project to rebuild it had been halted.

Revive the halted project and deliver a new pipeline that establishes clear ownership, serves all domains with GDPR-compliant access controls, and replaces the legacy pipeline.

Business case: Quantified the value of the new pipeline (including development velocity gains and cloud cost optimization) which brought the project back onto the roadmap.

Technical delivery: Architected and productionized the full pipeline: orchestration by Airflow, modeling by dbt with star schema (facts and dimensions), storage in BigQuery with partitioned tables, and documentation including comprehensive guides with examples for all users.

Pipeline productionized
Clear ownership structure established
GDPR-compliant access implemented
Documentation enabled smooth onboarding
Migration guide handed over to the legacy pipeline owners

Hand rehabilitation after injury or stroke requires repetitive, controlled movement therapy. Traditional devices are rigid, expensive, and often intimidating for patients. Soft robotics offers a gentler alternative, but designing a safe, controllable soft actuator for clinical use is a multidisciplinary challenge spanning mechanics, control theory, and fabrication.

Design, model, build, and validate a soft robotic hand rehabilitation device, from concept to working prototype.

Design: Designed the soft actuator geometry from scratch, balancing compliance for safety with force output for effective therapy.

Modeling: Built both analytical and numerical models to predict actuator behavior under pressure, enabling model-based controller design.

Control: Designed controllers to ensure safe, repeatable rehabilitation movements—automating the therapy process.

Fabrication: Molded and 3D-printed the device, set up and configured the pneumatic control setup, including real-time signal processing for pressure and position feedback.

Validation: Tested the full system experimentally to verify performance and safety.

Produced a fully functional soft robotic rehabilitation prototype
Demonstrated real-time viability in a peer-reviewed publication
Completed end-to-end delivery from conceptual design to real-time prototype testing

Automated emergency braking and steering systems often fail in sudden obstacle scenarios because they're too slow or too conservative. Testing these systems on real cars (not just simulations) is expensive, risky, and rarely done. I had the opportunity to work on this at e-Mobility & Control lab's electric vehicle lab.

Implement and test a collision avoidance system on a real electric vehicle. The goal was to prove that an intelligent control system can react faster and more safely than human drivers in "moose test" (sudden static obstacles on the road) scenarios .

On-the-fly adaptation: Upon starting the assignment, I discovered that the platform differed significantly from what we had assumed in the proposal. I rewrote the implementation plan on the spot to work with the actual hardware and sensors available.

Design & testing: Developed a dual-planner system combining predictive control with a human-like "maximum steering" fallback. Tested it across multiple speeds and hazard levels on a real test track and validated the system's ability to avoid collisions where the controller alone failed.

See it in action:

Stock claims were accumulating due to orders stuck in an undefined status. The root cause was unknown, the scope was unclear, and the issue was affecting multiple stakeholders.

Investigate the issue independently, identify the root cause, quantify the impact, and deliver a solution.

First, conducted a cross-functional investigation, gathering insights from multiple teams to piece together the full picture. Identified that orders which couldn't be fulfilled were later holding stock claims, taking up storage space and consuming purchasing capacity, even after refunds had been processed.

Second, designed and implemented an automated process to remove known stock claims. Performed statistical analysis to quantify the impact, establishing confidence intervals for cost recovery.

Root cause identified and documented
Automated stock claim removal flag implemented
Statistical analysis confirmed measurable cost recovery
Purchasing capacity improved

Automated driving systems often take overly conservative decisions in emergency situations. For example, they take unnecessary braking because they can't predict other drivers' behavior. This creates safety risks (e.g., rear-end collisions) and reduces trust in autonomous systems.

Develop a control strategy that enables proactive emergency maneuvers (such as safely changing lanes to avoid a collision) rather than just reacting with conservative braking.

Technical delivery: Designed a stochastic predictive planner that accounts for uncertainty in other vehicles' behavior. The design uses advanced approximation techniques to enable real-time implementation.

Verification: Validated the approach in industry standard high-fidelity simulations and benchmarked against state-of-the-art methods.

See it in action!

Real-time control systems in automated driving face a fundamental trade-off: accuracy vs. computational speed. Nonlinear control algorithms deliver better performance but are too slow for real-time applications, which is a critical barrier to commercialization.

Design a practical solution that makes nonlinear control computationally tractable without sacrificing performance that is accessible to engineers and academics.

Technical delivery: Developed H4MPC, an open-source MATLAB toolbox with a graphical user interface that automates the process of approximating complex nonlinear functions with computationally efficient alternatives.

Verification: Built two real-world case studies to validate performance: vehicle dynamics and tire saturation constraints.

Presented toolbox at IEEE Intelligent Vehicles Symposium, Kyoto, Japan
Open-source release with active adoption
Demonstrated real-time viability in a peer-reviewed publication
Bridged the applicability gap between academic research and commercial incorporation

The development team was building an automated system to merge orders placed in quick succession. The business rules had been defined, but they hadn't been tested against real-world data. Plus, the business case justifying the project was based on assumptions that needed verification.

Verify the business rules against actual order data, identify bottlenecks, validate or correct the business case, and present findings to the company.

Analyzed real order data to verify the business rules against actual patterns. Identified bottlenecks where the rules didn't match reality. Updated the business case with corrected assumptions and more accurate projections.

Business rules validated and corrected.
Business case updated with accurate assumptions and expectations.
Delivered company-wide demo, explaining the gap between expected and actual gains, and what we learned from the discrepancy
During the analysis, uncovered patterns that enabled the fraud detection team to identify previously unknown fraudulent cases.

A widely-used dashboard tracking customer behavior flow across an e-commerce website was broken. Built on GA4 data, the dashboard was displaying incorrect metrics due to problems in the pipeline. Multiple teams relied on this dashboard for decision-making.

Fix the dashboard by identifying what was broken in the query layer, YAML configs, and LookML. Deliver a working dashboard that stakeholders could trust.

Worked with a team of analysts to diagnose the root cause. I wrote complex SQL queries against GA4 data in BigQuery to validate what the dashboard was displaying vs. what the raw data actually contained. Moreover, rewrote LookML code to properly model the data for reporting, all while collaborated closely with stakeholders to understand their actual decision-making needs.

Dashboard accuracy restored through pipeline fixes
Delivered the project on time
Findings presented to stakeholders