Project showcase
Three core systems I built across R&D and Manufacturing Engineering, each from requirements to deployment and handoff, plus smaller yield, AI and characterization projects.
Optical characterization fixture
What
An automated 3-axis optical test fixture that maps the sensor's LED output in 3D with micrometer-level positioning accuracy, taken from design to a documented platform the team runs.
Why
The sensor's optical readings depend on exactly where the probe sits relative to the LEDs. We built this to identify the manufacturing tolerances for optical testing, the basis for test limits and hardware choices like glass versus plastic.
How
- Built motion control and data acquisition for a Thorlabs 3-axis stage (K-Cube controllers) positioning a fiber probe to micrometer accuracy, reading Feasa LED and IR analysers across configurable scan volumes and user-defined motion profiles
- Designed and toleranced the fixture, then ran repeatability and probe-tolerance studies on WHOOP 5.0 and next-generation hardware
- Wrote the LED calibration method and visualization tooling, documented SOPs, and onboarded engineers to run it independently
- Python
- NumPy
- pandas
- Matplotlib
- Plotly
- PyQt
- Streamlit
- Thorlabs K-Cube
- Feasa LED and IR analysers
- Fixture design and tolerancing
Inside the testbed
Seals out ambient light. A light-absorbing interior suppresses stray internal reflections, so the probe measures only direct LED emission.
Everything bolts to an optical breadboard. Three motor controllers drive the stage, one per axis.
Motorized X, Y and Z translation stages move the probe through space above the device.
A dowel-pinned aluminum bracket on the Z stage carries the fiber probe head. A pocket in the base plate locates the device.
Fiber carries the collected light to Feasa LED and IR analysers that read intensity and color.
Sensor side up in its pocket. The LEDs switch on and the probe maps the light above them.
Results
- Identified the manufacturing tolerances for optical testing, with repeatability limits and probe tolerances for WHOOP 5.0 and next-generation hardware
- Created a hardware abstraction layer and simulation software, reused across other projects
- Documented platform handed off; engineers now run it independently
On-device capacitor triage system
What
A production screen that flags failing battery capacitors from their current signature, running at the contract manufacturers during final assembly.
Why
A battery capacitor failure was reaching the field. Nothing at the time could reliably identify it in a short enough time for production; state-of-charge and voltage tests both fell short.
How
- Wrote device firmware in C to add raw battery-current streaming over BLE, then extracted FFT features for an inline logistic-regression classifier
- Validated against failure-analysis teardowns, CT-scan correlation and a large production DOE run with the contract manufacturer
- Built the full-stack monitoring platform (Python backend, PyQt and Streamlit front ends), deployed locally at the contract manufacturers during final assembly and integrated into the production test system behind a feature flag
- C
- Embedded firmware
- BLE
- Python
- NumPy
- SciPy
- scikit-learn
- pandas
- PyQt
- Streamlit
- FFT
- Machine learning
- DOE
Current spectrum of one unit. Illustrative model, not WHOOP data.
Results
- $1.8M saved on device returns and logistics costs
- 99% accuracy flagging bad capacitors
- Screen cut from 250 s to about 20 s
- Shipped to production at the contract manufacturers
ATE test software
What
A distributed hardware-in-the-loop test framework that runs independent test steps side by side across a fleet of fixtures.
Why
Legacy ATE ran every test step in series, so fleet run time grew with every step and fixture added.
How
- Re-architected serial legacy ATE into dependency-aware, DAG-scheduled test orchestration in Python across a fleet of fixtures
- Raised pytest coverage to 95% and used fleet-wide analysis to separate rig faults from real device defects
- Prototyped and demoed the DAG orchestration layer for next-generation test software
- Python
- pytest
- DAG orchestration
- Hardware-in-the-loop
Ready
Click any step to make it fail on the next run. Illustrative steps and timings.
Results
- 30% faster test time
- 95% pytest coverage
- 40% fewer false failures
Additional projects
Sensor yield improvement
Root-caused IR LED measurement noise to test-software timing in the analog front end across 10+ experiments, then shipped the fix to manufacturing software.
- Root-cause analysis
- GRR studies
- AFE timing
Failure Analysis Bot
AI failure-analysis assistant over the first consolidated failure-analysis database. Benchmarked and evaluated sentence-transformer models on manufacturing-specific retrieval, served semantic search from local Hugging Face models with embeddings in Postgres, and used OpenAI to fill the gaps. Piloted in manufacturing.
- Hugging Face sentence transformers
- PostgreSQL
- OpenAI API
- RAG
- Model benchmarks and evals
Optical stack MFG testing
Devised manufacturing test methods for dark noise, black-box, and LED and photodiode assessments at board and assembly level on the manufacturing line.
- Manufacturing test design
- Dark noise
- Black-box testing
- LED and photodiode test
IMU QC fixture
Developed software to monitor fixture reference IMU drift over time, and built benchtop fixtures to recalibrate them.
- I2C
- Drift monitoring
- Fixture design
JTAG research
Conducted extensive JTAG boundary-scan implementation research on current and NPI hardware, coordinating with suppliers and contract manufacturers.
- JTAG
- FPGAs
- International communication