Optical engineer and data scientist · PhD in nano-optics
I turn measurements and records into tested, explainable models: from photonic devices to patient data.

punjdeep.com · LinkedIn · [email protected] · Ghent, Belgium

Ten years of optics research at Institut Fresnel, KAUST, UC Irvine, IIT Madras and Leiden University, and now at Ghent University's Faculty of Pharmaceutical Sciences, where I build optical set-ups and the Python behind them: instrument control, image analysis and quality checks of laser systems. First-author paper in Nature Nanotechnology; h-index 13.

On GitHub I build complete projects in two areas: data science on healthcare, industrial and energy data, and photonics design and packaging.

Open to data science and ML roles (healthcare, pharma, industry, energy) and optical / photonics engineering roles.

Patient finding: undiagnosed kidney disease · healthcare
Which adults will be diagnosed with chronic kidney disease in the next 12 months?

• Yearly patient snapshots built with SQL in PostgreSQL from 10,000 synthetic health records; features only from the 2 years before each snapshot, checked by leakage tests
• Rare target (0.5% positive): models compared on PR-AUC, trained on 2019-2023 and tested on 2024-2025, runs tracked in MLflow
• Result: reviewing the top 10% of patients finds 49% of next year's cases; 42% of new patients had an abnormal kidney test a median of 55 months before diagnosis

Python PostgreSQL scikit-learn MLflow Streamlit pytest

Laser weld defect inspection · industrial ML, thermal imaging
Can a thermal camera tell when a laser weld turns defective?

• 16,362 labelled infrared frames from 97 real welds; 7 physics-based features computed without labels or future frames
• Random Forest with nested cross-validation split by weld, checked against a leakage baseline; SHAP explanations
• Result: ROC AUC 0.974; 91% of defective frames detected at 5.6% false alarms

Python scikit-learn SHAP Streamlit pytest GitHub Actions

Battery dispatch on Belgian electricity prices · energy, forecasting + optimisation
How much can a battery earn when it is scheduled from price forecasts instead of known prices?

• Day-ahead price forecast (gradient boosting) on 10 years of hourly prices, chronological train / validation / test split
• Daily battery schedule as a linear program (CVXPY), made from forecasts only and settled at real 2024 prices
• Result: forecast error 17.2 vs 23.2 EUR/MWh for the same-hour-yesterday baseline; 73.5% of the perfect-foresight profit, only about 1 point above the baseline

Python pandas scikit-learn CVXPY

SiN edge-coupler study · silicon photonics, packaging
From waveguide mode design to assembly yield for fibre-to-chip and chip-to-chip coupling at 1550 nm.

• Finite-difference mode solver checked against the exact slab solution; misalignment, gap loss and Monte Carlo assembly yield
• Gaussian-process surrogate for the process window; expanded-beam micro-optic relay ray-traced in Optiland; test-chip DOE layout in GDS
• Result: 0.43 dB coupling to UHNA7 fibre; at a 1 dB spec, 96% assembly yield at 0.25 µm placement accuracy vs 19% at 1 µm

Python NumPy/SciPy Optiland gdstk pytest GitHub Actions

Last update: fab-reporting-ba, 2026-10-26

US drug shortages dashboard
Power BI dashboard analyzing 1,141 active shortage listings (69 drugs) by category, manufacturer and duration

data-analysis data-visualization powerbi

Earlier data projects
Earlier projects: Goodreads Streamlit app, Tableau dashboards, breast-cancer and banknote classifiers

portfolio-project scikit-learn streamlit tableau

Fab reporting ba
Business analysis case for a semiconductor pilot line: requirements, KPI dictionary, Kimball star schema, data-quality checks, SQL analysis, Streamlit dashboard and Power BI model. Finds the litho bottleneck and an etch-tool yield excursion. Simulated data.

business-analysis business-intelligence data-analytics data-modeling dax kimball kpi-dashboard manufacturing-analytics pandas portfolio-project postgresql power-bi python semiconductor sql star-schema streamlit yield-analysis

• Evaluation that can't cheat: data split by time (patients, prices) or by group (welds), and always compared with a simple baseline or an exact result
• Reproducible: one environment file per project, fixed seeds, data downloaded or generated by script
• Tested: pytest checks for leakage and physics, run on GitHub Actions where the project allows it
• Written up honestly: every README states the result, how it was measured and its limitations