Swapneel Wagholikar

Defense and public-safety imaging systems. Described at the level of method and problem class; program details omitted.

Airborne perception and system integration

• Co-designed an embedded airborne imaging system — compute and sensor selection, and the interfaces carrying commands, telemetry and video — then trained the object detection model that runs on it, deployed as a TensorRT engine on NVIDIA Jetson with detections geo-located against terrain elevation data.

Infrared calibration and image quality

• Own the production calibration automation for uncooled LWIR camera cores, and characterized focal-plane noise with the NVESD 3D noise model to decide which denoising blocks — destriping, spatial and temporal denoise, flare correction — were worth writing.

ISP block design

• Designed the auto-exposure, histogram tone mapping and autofocus blocks for a broadband SWIR/MWIR camera from scratch, then wrote the bit-accurate fixed-point golden reference model the FPGA team verified their RTL against.

Pointing, orientation and visual-inertial estimation

• Rewrote the celestial calibration math behind a heading reference, and built a drift-free orientation reference by calibrating pan-tilt encoders against an IMU through quaternion mappings, boresighted onto the optical axis.

Geospatial segmentation and route planning

• Built satellite-imagery semantic segmentation end to end and the 2.5D route planner on top of it, argued that planner onto a classical algorithm for lack of training data, and ported it to C++ for ATAK integration.

Monocular depth on edge hardware

• Built monocular depth estimation for a mobile vending robot with a MiDaS-family model, optimized with TensorRT for real-time inference on a Jetson Nano.

LiDAR-camera fusion and calibration

• Wrote a C++/ROS pipeline that synchronises a Velodyne solid-state LiDAR with an Intel RealSense stereo camera inside a 50 ms window, projecting the two into LiDAR-backed ground-truth depth maps, with auto-calibration recovering the extrinsics so no manual alignment pass was needed.

Depth completion and edge deployment

• Built a depth-completion pipeline on CompletionFormer that turns sparse LiDAR returns into dense depth — fine-tuning and synthetic augmentation cut RMSE 81% and MAE 92% against the pretrained model — then quantized it to INT8 for edge deployment.

Retrieval-augmented LLM interface

• Built an LLM interface that lets business users query and visualise economic records, using retrieval-augmented generation over OpenHermes-2.5 and the vector store behind it.

Economic forecasting

• Ran rolling-window forecasts across 50 years of 118 economic indicators with LSTM and VARIMA models, reaching 0.04 MAPE, and estimated recession probability with a support vector machine over the same set.

GPS, visual and inertial odometry

• Fused RTK-GPS, a ZED2 stereo camera and an IMU into an odometry module holding position within a centimetre in motion, under open sky with an RTK fix.

SLAM packaging

• Packaged an error-state Kalman filter for SLAM as a Dockerised ROS2 node.

Transformer point-cloud completion

• Trained and extended PointAttN, a transformer for point-cloud completion under Prof. Ziming Zhang, adding a hyper-spherical module after the encoder that improved chamfer distance 33% over the published baseline to 0.0929 mm.

Mechanical design and documentation

• Produced technical packages for oil-free compressor designs covering 3D models, P&ID and bill of materials, reviewed designs through the internal audit process, and mentored junior engineers on design standards.

GPA 4.0/4.0. Coursework across computer vision, deep learning, reinforcement learning, motion planning and robot control.

CGPA 9.32/10.