Projects — Mohd Bilal

Finite Element Analysis of Proximity-Induced Interaction Between Horseshoe-Shaped and Circular TBM Tunnels in Soft Ground In Progress

Collaborators: Safoora Hafeez · Dr. Md. Rehan Sadique · Dr. Zaid Mohammad
Tools: OPTUM G2 · FEM Static Analysis · Geotechnical Modelling

Conducted a comprehensive 2D Finite Element Method (FEM) static analysis in OPTUM G2 to evaluate the proximity-induced soil-structure interaction between an existing horseshoe-shaped tunnel and a newly excavated circular TBM tunnel in soft ground.
Systematically evaluated sequential excavation across four center-to-center spacing configurations (2.5D, 2.0D, 1.5D, and 1.0D) to measure ground deformation, lining convergence, and internal stress distribution.
Identified asymmetric deformation modes, ovaling, and elevated lateral convergence in the horseshoe tunnel under close spacing, alongside deepening surface settlement troughs caused by shear strain localization and diminished soil arching.

Prediction of Foundation Settlement Using Hybrid FEA–ML Framework

Tools: OptumG2 · Pandas · scikit-learn · XGBoost

Performed 50+ finite element simulations in OptumG2 to model immediate settlement of strip footings under drained soil conditions with varying parameters.
Trained an XGBoost-based ML model for predicting immediate settlement of shallow foundations using FEA-generated data, achieving R² > 0.92 and mean absolute error < 8%.
Validated model predictions against real field case-history data, demonstrating strong agreement with observed behavior.

GIS-Based Spatial Analysis for Infrastructure Planning

Tools: QGIS · Python · GeoPandas · Rasterio · OpenStreetMap

Conducted GIS-based multi-criteria spatial analysis over a 100–200 km² study area to identify optimal zones for infrastructure development.
Generated weighted suitability maps integrating terrain slope, soil constraints, land-use restrictions, and road proximity, reducing unsuitable land identification by 40%.

Nuclear Power Plant Site Suitability & Hazard Mapping GitHub ↗

Tools: Python · QGIS · scikit-learn · Random Forest · GeoPandas · K-Means · GeoTIFF

Built a geospatial ML pipeline that answers two questions simultaneously: where should new nuclear plants be sited, and how hazardous are existing ones?
Trained a Random Forest model on existing plant locations to produce a continuous 0–1 suitability raster (GeoTIFF) scoring every pixel in the study region across geological, demographic, and infrastructure criteria.
Generated wind-adjusted IAEA emergency zones per plant (PAZ 5 km / UPZ 30 km / LCPZ 300 km) with population exposure counts and K-Means risk tiering (Low / Medium / High / Critical), combined into a single interactive browser map.

AUV Localization and Control System

Tools: PyTorch · OpenCV · IMU Sensors · PID Control

Developed an Autonomous Underwater Vehicle (AUV) localization and control system integrating IMU data with control feedback for motion estimation.
Implemented PID-based depth control for stable underwater navigation in low-visibility conditions.
Trained and integrated a custom SuperPoint detection model with visual–inertial cues to improve localization and pose estimation.

ASV Waypoint Navigation System — ANDROMEIDA GitHub ↗

Collaborators: Tabish Shah Mohsin · Mohammad Ayan
Tools: Python · MAVLink · Pixhawk · Jetson Orin Nano · Leaflet.js · UDP · PID Control · RTK GPS · VectorNav VN-100 · PNI RM3100

Designed and built a full navigation stack for an Autonomous Surface Vehicle using a dual-compute architecture — a Jetson Orin Nano onboard and a Mac-based Ground Control Station communicating over UDP (ports 5005/5006).
Built a web-based GCS (Python HTTPServer + Leaflet.js) with live Chart.js telemetry graphs, GPS-based map initialization, and MAVLink integration for Pixhawk 2.4.8 sensor data (GPS_RAW_INT, attitude, IMU).
Implemented PID-based yaw control using a complementary heading filter combining RM3100 absolute magnetic heading and VN-100 yaw rate; binary UART packets with CRC validation relay thrust commands to the STM32H755 ESC controller with a hardware watchdog failsafe.

DevFlow AI — Agentic Developer Productivity Assistant GitHub ↗

Tools: Python · FastAPI · LangChain · Ollama (LLaMA 3.1:8b) · Google Calendar API · Google Tasks API · GitHub API · Langfuse · MCP

Built an agentic AI assistant that lets developers manage tasks, schedule work sessions, and interact with GitHub repositories through natural language — no manual tool switching required.
Architected a modular system with a FastAPI REST backend, a LangChain reasoning agent, and an MCP Tool Orchestrator coordinating calls to Google Calendar, Google Tasks, and GitHub APIs from a single conversation interface.
Integrated Langfuse observability for full agent tracing, metrics collection, and performance monitoring; designed with extensibility for Jira, Linear, Slack, and CI/CD platforms.
Runs on a local LLM (LLaMA 3.1:8b via Ollama) for privacy-first deployment with no dependency on external AI APIs.

Traffic Sign Recognition — CNN vs Random Forest Benchmark GitHub ↗

Tools: Python · TensorFlow · scikit-learn · HOG Feature Extraction · GTSRB Dataset

Built and benchmarked two competing approaches for traffic sign recognition on the German Traffic Sign Recognition Benchmark (GTSRB) — 43 classes, images resized to 64×64.
CNN (TensorFlow/Keras): convolutional layers with ReLU activations, MaxPooling, Dropout regularization, and fully connected classification head — achieved 94.76% test accuracy.
Random Forest (scikit-learn): HOG (Histogram of Oriented Gradients) feature extraction with hyperparameter-optimized ensemble — achieved 89.68% test accuracy with interpretable feature importance analysis.
Produced detailed confusion matrices and training metric plots for both models, surfacing class-level misclassification patterns and guiding decisions on data augmentation.

Technoxian Competition Robots — Line Follower & RoboHockey GitHub ↗

Collaborators: Mohammad Shiraz · Supreet Chaudhary · Kandarp Gupta
Tools: C++ · Arduino · PID Control · IR Sensors

Wrote Arduino firmware in C++ for two competition robots entered in the Technoxian robotics competition: a PID-controlled line follower and a RoboHockey bot.
Implemented sensor-driven control loops using IR arrays for line detection and tuned PID gains for smooth, high-speed line tracking without oscillation.
Served as an early embedded systems project establishing foundational skills in real-time control, motor drivers, and hardware debugging that carried forward into AUV and ASV work.

SeaThru-NeRF — Underwater 3D Scene Reconstruction GitHub ↗

Tools: Python · Neural Radiance Fields · COLMAP · PyTorch

Implemented a SeaThru-NeRF model for underwater 3D scene reconstruction, extending standard NeRF with a medium-aware scattering model to reverse the degrading effects of water (backscatter, attenuation, color shift).
Built a COLMAP-to-NeRF conversion pipeline (convert_colmap.py) for processing real underwater image captures into training-ready camera pose data.
Motivated by AUV perception work — accurate 3D mapping of underwater environments is directly applicable to autonomous navigation and obstacle avoidance in low-visibility conditions.

ROV 4.0 — Remotely Operated Underwater Vehicle GitHub ↗

Tools: Python · OpenCV · UDP · Multithreading · Raspberry Pi · Tkinter

Co-developed the software system for ZHCET AMU's MTS AUV Club ROV, structured around a Vehicle Side and Base Station Side communicating over UDP.
Vehicle side uses a multithreaded producer-consumer architecture: a run() producer thread generates thruster PWM control values and pushes them onto a thread-safe queue, while a GUI() consumer thread encodes and transmits data back to base station.
Implemented a separate live camera stream using OpenCV over UDP, with base station scripts for rendering telemetry and video feed; designed USB/IP forwarding setup for remote joystick input from base station to vehicle over the network.