Selected work

SELECTED
WORK.

05 Projects
2024—26 Timeline
Scroll to explore
01 / 05
Astrodynamics · Machine Learning · HMI

Orbital Mechanics
Tools & AI Surrogate
Dashboard

Personal Project · Mar – Apr 2026

Full-stack Python astrodynamics suite — Hohmann, Bi-Elliptic & Phasing transfers computed analytically, paired with an MLPRegressor surrogate achieving R² = 0.9993 on 40,000 held-out mission scenarios. PySide6 HMI embeds live Matplotlib orbit visualisations; covers Earth, Moon, and Mars.

R² Score
0.9993
MAPE
2.88 %
Training scenarios
200 k
Python PySide6 scikit-learn NumPy Matplotlib Joblib
GitHub View Details →
Phase 01 — Live Dashboard
Orbital Transfer Calculator Dashboard
Live Dashboard — Orbital Transfer Calculator
PySide6 HMI showing a Hohmann transfer from LEO to GEO. Input panel on the left, physics results and AI surrogate prediction side-by-side on the right, with an embedded Matplotlib orbit plot drawn to scale at the bottom.
Physics Engine
Hohmann Transfer
Bi-Elliptic Transfer
Phasing Manoeuvres
Tsiolkovsky Rocket Eq.
AI Surrogate
R² = 0.9993
MAE = 16.47 m/s
MAPE = 2.88 %
40,000 test scenarios
Synthetic dataset generation — 200,000 mission scenarios
generate_data.py — 200,000 Mission Scenarios
Terminal output of the synthetic data pipeline: orbit radii, spacecraft mass, Isp, and manoeuvre type are randomly sampled; the physics engine computes the exact delta-v and propellant mass ground truth for each scenario. The result is saved as orbital_data.csv.
Training Coverage
Earth · Moon · Mars
LEO → Trans-Lunar
v_c: 261 – 7,784 m/s
No body flags needed
Dataset Split
200,000 total
160,000 train (80 %)
40,000 test (20 %)
Clean, zero measurement noise
Neural network training pipeline — 5-stage MLP
train_model.py — 5-Stage MLP Training Pipeline
The terminal logs all five stages: dataset load → StandardScaler preprocessing → 80/20 split → MLPRegressor training → evaluation on 40,000 held-out samples. Final output: R² = 0.9993, MAE = 16.47 m/s, RMSE = 30.56 m/s, MAPE = 2.88 %. Both model and scaler serialised with Joblib.
Model Metrics
MAE = 16.47 m/s
RMSE = 30.56 m/s
MAPE = 2.88 %
R² = 0.9993
Stack
scikit-learn MLPRegressor
StandardScaler
Pandas · NumPy
Joblib serialisation
01
02 / 05
Machine Learning · Aerospace

Prognostics & Health
Management System
for Gas Turbine Engine

Personal Project · Oct – Dec 2025

End-to-end ML pipeline predicting engine Remaining Useful Life on the NASA C-MAPSS dataset. Custom Asymmetric Loss Function cuts safety violations from 12.4 % to 1.09 %. LSTM outperforms Random Forest baseline (RMSE 17.17 → 13.32 cycles).

RMSE Reduction
78 %
Safety Violations
↓ 91 %
Model Accuracy
87 %
LSTM Python MLflow DVC K-Means NASA C-MAPSS
GitHub View Details →
Phase 01 — Exploratory Data Analysis
Sensor 7 Degradation Trend
Sensor 7 Degradation Trend
The first proof: plotting sensor values against time-in-cycles reveals a clear noisy downward trend across all engines — the degradation signal the model must learn.
Feature Correlation with RUL
Feature Correlation with RUL
All 21 sensors ranked by predictive power. Top predictors (sensor_12, sensor_7) are physically linked to the HPC fault mode — the exact failure in this dataset. Six zero-correlation sensors dropped.
RUL vs Sensor 7 Relationship
RUL vs. Sensor 7
Piecewise linear scatter confirms the data engineering strategy: clipping RUL at 125 cycles simplifies the healthy phase, letting the model focus on the critical degradation zone.
RUL vs Sensor 12 Relationship
RUL vs. Sensor 12
Sensor 12 mirrors sensor 7 — both are HPC-linked indicators. Their near-identical shapes reveal strong multicollinearity: no issue for LSTM, but fatal for linear models.
Scatter Matrix of Key Features vs RUL
Scatter Matrix — Important Features vs. RUL
The full picture: top sensors are tightly correlated with each other and with RUL. This multicollinearity rules out linear regression and motivates the LSTM + cluster similarity approach.
K-Means Cluster Similarity
K-Means Operating Modes
Engines behave differently at different altitudes and throttles. K-Means on sensors 7 & 12 discovers these regimes automatically. Each point's similarity to cluster centers (X) becomes a "Health Signature" feature.
ColumnTransformer Pipeline Architecture
Master Pipeline — ColumnTransformer
The full preprocessing architecture: log-transforms for exponential decay sensors, K-Means similarity injection, and automatic unit ID stripping. A Group Split by Engine ID prevents data leakage across time.
Linear Regression Baseline Performance
Linear Regression — RMSE 17.17
The baseline "smear": predictions scatter widely around the perfect line. Engine degradation is exponential and non-linear — a straight line cannot capture the failure curve.
Random Forest Baseline Performance
Random Forest — RMSE 16.05
A tighter fit to the perfect prediction line, but still with overgeneralisation at the extremes. Random Forest lacks "memory" — it sees one row at a time with no sense of degradation velocity.
LSTM Model Architecture
LSTM Architecture
Two stacked LSTM layers (64 → 32 units) with Dropout(0.2). Input: (Samples, 30 TimeSteps, Features) — feeding the last 30 flight cycles so the model learns velocity and acceleration of degradation, not just absolute values.
LSTM Training History
LSTM Training History
Training and validation loss converge cleanly. No divergence between curves confirms the Dropout regularisation is working — the model generalises rather than memorising.
Random Forest Safety Analysis — Baseline
RF Safety Baseline — 12.41% Risk
Safety analysis of the Random Forest baseline. A tail of predictions falls in the Dangerous Zone (overestimation). A 1-in-8 chance of predicting a safe engine that is actually failing is unacceptable for safety-critical operations.
LSTM Error Distribution 10.26% Risk
LSTM Error Distribution — 10.26% Risk
The standard LSTM reduces dangerous overestimations from 12.41% to 10.26% — a genuine improvement. But 1 in 10 is still too high. Accuracy alone is not enough for safety-critical maintenance.
Random Forest Feature Importance
Random Forest Feature Importance
The K-Means cluster similarity features dominate importance — validating the pipeline design decision. "Cluster_0_similarity" alone accounts for ~37% of predictive power, confirming operating-mode awareness is the key engineered signal.
Keras Tuner Hyperparameter Distributions
Keras Tuner — Hyperparameter Search
Keras Tuner mathematically optimises the LSTM architecture across layers, units, and dropout. Loguniform distributions efficiently explore learning rates across orders of magnitude, bringing RMSE to 13.32 cycles.
Tuned LSTM Training History
Tuned LSTM Training History
The optimised model converges faster and to a lower loss floor. Training and validation curves remain tightly coupled across all 22 epochs — no overfitting despite the increased model capacity.
Tuned LSTM Performance RMSE 13.32
Tuned LSTM — RMSE 13.32
The tuned model is more accurate than the baseline LSTM, but standard MSE loss still treats overestimation and underestimation equally. The risk factor stays at 9.60% — accuracy and safety are separate problems.
Asymmetric Loss Impact — Risk 1.09%
Asymmetric Loss — Risk Drops to 1.09%
A custom TensorFlow loss applies a 10× penalty to overestimation errors. Gradient Descent is mathematically forced to "fear the danger zone." Risk collapses from 9.60% → 1.09% — a safety breakthrough.
Safety-First Model Scatter
Safety-First Model — Conservative Estimates
Green predictions have shifted below the diagonal — the model now systematically errs on the side of caution. Maintenance crews will always be alerted before a failure becomes critical, not after.
Engine 1 Lifecycle Tracking
Engine #1 — Full Lifecycle Tracked
The AI (green) tracks the ground truth (black dashed) closely across all 160 flight cycles, but stays slightly below — safe conservatism in action. This is the ideal human-machine interaction signature for predictive maintenance.
Final Test Set Performance RMSE 18.79
Final Exam — Unseen Test Engines, RMSE 18.79
Testing on a completely held-out set of engines never seen during training or validation. RMSE of 18.79 cycles reflects real-world generalisation cost — but the safety profile holds: predictions consistently avoid dangerous overestimation zones.
02
Section 01 — Simulator Rig
Blurred images are reference only — project is under NDA
Developer in the simulator rig
Inside the Simulator Dome — FSR TU Darmstadt
All three monitors running the PySide6 GUI live in the FSR dome. Left: Flight Controls Panel; centre: aircraft synoptic; right: backup systems. Every panel driven by real X-Plane 12 telemetry streamed over UDP at 50 Hz.
Stack
PySide6 · X-Plane 12
XHSI Plugin · UDP / Multi-threading
Python 3 · QTimer sync
Key Numbers
50 Hz data rate
< 20 ms end-to-end latency
Grade 1.3 awarded
PFD and MFD Display
Primary Flight Display + MFD — XHSI Plugin
Left: ND (Navigation Display / HSI) with heading rose and traffic advisory. Centre: PFD with attitude indicator, speed tape, altitude tape, and FD bars. Right: dual-engine instruments. All rendered via the XHSI (X-Plane HSI) 3rd-party plugin, fed by live X-Plane 12 datarefs over UDP.
Live flight test with GUI active
GUI Active — Live Flight Test
All panels updating simultaneously during an active sim session. Demonstrates the sub-20 ms latency from X-Plane dataref change to on-screen display refresh across every concurrent widget.
Single Pilot Cockpit Overview — NDA blurred
Cockpit Layout Overview [NDA]
Full single-pilot window arrangement: ND + PFD + engine column on top, additional systems and synoptic pages below. Multiple independent PySide6 widgets tiled into one unified operator view.
HYD/FUEL and ELEC/AIR systems live testing
HYD & Fuel / ELEC & Air — Live Test
Bleed air system page streaming live from X-Plane. ENG 1/APU/ENG 2 BLEED states, XFEED AUTO, and HOT AIR buttons all respond directly to X-Plane datarefs over the multi-threaded UDP connection.
ECAM Systems Pages — NDA blurred
ECAM Systems Pages Grid [NDA]
Four synoptic pages displayed at once: electrical bus diagram, system data table, hydraulics schematic, and flight controls surface page with blue/yellow/green actuation colour-coding. Each page is a separate PySide6 widget on its own dataref set.
Flight Controls Panel UI with source code
Flight Controls Panel — PySide6 + Source
Side-by-side: FCP UI and FCPmain.py. Handles elevator trim (±3° animated slider), rudder trim, gear & brakes, speed brakes, and flaps — each writing back to X-Plane datarefs. QTimer syncs control state from the sim every 30 s on startup.
Flight Controls Panel — NDA blurred
FCP Operational View [NDA]
Elevator trim indicator, gear & brakes lights, speed brake and flap sliders (Low / Med / Max). All controls bi-directionally linked — physical inceptor movement in X-Plane updates the GUI in real time.
ATC and Backup Controls — NDA blurred
ATC + Backup Controls [NDA]
ATC radio management: active/standby frequency, VHF1/VHF2 select, and live command-history log. Backup Controls provide a fallback interface if primary displays fail — fulfilling the single-pilot redundancy requirement.
Nose light slider widget
Nose Light — AnimatedSlider
Custom PySide6 AnimatedSlider with discrete T.O / TAXI detents for nose landing-light control.
Reusable Component
The same AnimatedSlider class drives speed brakes, flap levers, and all other slider-based controls — write once, deploy everywhere across the GUI.
03 / 05
HMI · Aerospace · Software

Single-Pilot Simulator
Displays

FSR — TU Darmstadt · Apr – Sep 2025

Modular GUI for single-pilot cockpit operations built with PySide6. Multi-threaded UDP layer streams real-time telemetry from X-Plane 12 at 50 Hz with sub-20 ms latency. Awarded a 1.3 grade for HMI performance and backend reliability.

Data Rate
50 Hz
Grade Score
1.3
Latency
< 20 ms
PySide6 X-Plane SDK UDP Multi-threading HMI
GitHub View Details →
03
04 / 05
Safety Analysis · Aerospace · ML

Runway Incursion
Safety Analysis

Boeing × FSR · Oct 2024 – Mar 2025

30 years of major runway incursion events analysed, focused on collision-critical scenarios. Sensitivity analysis conducted under the SURFIA framework and structured against RTCA DO-323. Findings presented to Boeing safety engineers and the FSR department at TU Darmstadt. Awarded a 1.7 grade.

Data Coverage
30 yrs
DO-323 Coverage
94 %
Grade Score
1.7
Safety Analysis SURFIA RTCA DO-323 Pandas NumPy Boeing
View Details →
Section 01 — Boeing Meeting & Dataset
Data visuals are intentionally obscured — project is under NDA
Boeing meeting visitor badge
Boeing Meeting — 06 Mar 2025
Visitor badge for the FSR × Boeing safety review at Boeing facilities. Runway incursion sensitivity analysis findings presented directly to Boeing safety engineers.
Sample of RI dataset — NDA protected
30-Year RI Dataset [NDA]
Sample of the runway incursion event database: date, airport, aircraft category, severity, contributing factors, and narrative fields. Major collision-critical events isolated from 30 years of ASRS / FAA records.
Dataset Scope
30 years of major RI events
Collision-critical scenarios only
ASRS + FAA incident records
Multi-column feature extraction
Delivered To
Boeing safety engineers
FSR — TU Darmstadt dept.
Grade awarded: 1.7
Framework
SURFIA
Surface Risk Identification & Assessment
Sensitivity analysis performed under the SURFIA methodology — a structured framework for identifying which contributing factors most influence runway incursion risk. Permutation-importance scoring ranked each feature by its marginal effect on collision probability, isolating the highest-impact variables for Boeing's safety model.
Standard
RTCA DO-323
Runway Incursion Safety — Operational Requirements
The entire safety logic and feature classification structure was mapped against RTCA DO-323 — the industry standard governing runway incursion avoidance system requirements. 94 % of the document's operational requirement categories were addressed in the analysis, ensuring the findings were directly actionable for Boeing's certification pipeline.
Analysis Pipeline
Data ingestion → cleaning
Feature engineering
SURFIA sensitivity scoring
DO-323 category mapping
Permutation importance
Boeing-ready report output
Tech Stack
Python · Pandas · NumPy
Scikit-learn
Permutation importance
RTCA DO-323 mapping
SURFIA framework
04
Phase 01 — Design
CAD design of height-adjustable mower unit
CAD Design — Fusion 360
Lead-screw driven height-adjustment mechanism with stepper motor mount and dual guide rails. Designed for full integration with the mowing deck assembly.
Simulink PID control model
Simulink — PID Control Model
Full closed-loop simulation: PID controller driving a stepper motor model with position, velocity, and acceleration feedback. Gains tuned in simulation before hardware deployment.
3D printing prototype parts on Prusa i3 MK3
Prototype — Prusa i3 MK3 Print
Physical prototype components printed on a Prusa i3 MK3. Motor mounts and structural brackets iterated through multiple print cycles until all stability specs were met on first hardware test.
05 / 05
Mechatronics · Control Systems

Height-Adjustable
Mower Unit

TU Darmstadt · Apr – Sep 2024

Active stabilisation control logic for a mechatronic mowing system using MATLAB/Simulink. Physical prototypes designed in Fusion 360 and 3D-printed. System integration managed via agile SCRUM. All stability specs met on first hardware test.

Stability Gain
88 %
Settle Time
↓ 75 %
Overshoot
< 5 %
MATLAB Simulink Fusion 360 SCRUM 3D Printing
View Details →
05
What's next

Currently
In Development

More projects dropping soon — check back or get in touch

← Back to Home Get in Touch