Filter: Identity
Layer: Input (1x1x3)

Mohanen Balasubramanian

Deep Learning
Computer Vision
AI Systems

Senior Lead Engineer

Passionate about advancing AI through Deep Learning and Computer Vision. Specializing in neural networks, multi-task learning, and real-time perception systems for autonomous driving.

Params: 0
Output: RGB Profile
Filter: Temporal
Layer: Conv3D (7x7x8)

Professional Experience

Valeo, Chennai, India

FEB 2021 - PRESENT

Senior Lead Engineer (AI R&D)

  • Multi-Task BEV Network (2025): Developed and optimized a unified neural network with shared encoder and multiple decoders for parametric lane and object detection; led the implementation of parametric lane detection from scratch and multi-task training strategy.
  • Lane Detection (2024): Built a point-based lane detection algorithm in Bird's Eye View using surround-view camera projections; implemented core BEV lane detection modules from the ground up.
  • Multi-Task Learning (2023): Contributed to a multi-task model supporting soiling detection, segmentation, object detection, and more; focused on trailer tracking and emergency vehicle detection decoders.
  • Targetless Trailer Tracking (2022): Customized a vision + kinematics fusion algorithm with UKF for trailer tracking to meet BMW's ADAS architecture needs.
  • End of Line Calibration (2021): Engineered feature detection and cost functions for custom target-based camera calibration, minimizing reprojection error without traditional checkerboards.
  • Motion Tracking Calibration (2021): Maintained and debugged dynamic camera extrinsics estimation using motion vector-based cost optimization for surround-view systems.

E-con Systems / VisAi Labs, Chennai, India

MAY 2016 - JAN 2021

Project Engineer

  • Individual Contributor.
  • Developed algorithms for computer vision applications (stereo depth, calibration, 2D processing, feature detection, 3D reconstruction, noise removal, clustering, measurements).
  • Developed products from scratch to production, with hands-on experience in system design and development.
  • Mentored team members and planned tasks for developed products.
Params: 8.5M
Output: Career Timeline
Filter: Feature
Layer: Pooling (2x2)

Key Projects

Multi Task BEV Network

Valeo (Driving Perception Platform) | 2025
  • Multi-task BEV model for parametric Lane detection and Object detection.
  • Single encoder, multiple decoders for various tasks.
  • Created and improved Parametric Lane Detection implementation from scratch and Multi-task training.

Lane Detection

Valeo (Driving Perception R&D) | 2024 - 2025
  • Point-based Birds Eye View Lane Detection.
  • Projects features from 4 surround-view cameras to BEV grid for lane detection.
  • Created and improved Point-based Lane Detection implementation from scratch.

Multi Task Learning

Valeo (BMW, Stellantis) | 2023 - 2024
  • Multi-task model: soiling detection, segmentation, object detection, privacy protection, trailer tracking.
  • Single encoder, multiple decoders.
  • Worked on Trailer tracking and Emergency Vehicle detection decoders.

Targetless Trailer Tracking

Valeo (GM, BMW) | 2021 - 2022
  • Tracks trailer for ADAS (jackknife prevention, automated parking/reverse).
  • Vision + kinematics fusion with Unscented Kalman Filter (UKF).
  • Customized algorithm architecture for BMW requirements.

vMeasure - Object Dimensioning Solution

Vis-Ai Labs | 2019 - 2020
  • Estimates 3D object dimensions (L, B, H).
  • Uses stereo cameras for depth -> point clouds -> noise removal -> custom min bounding box.
  • Developed from scratch; involved in system design, engineering, KPI creation, testing, estimation.

End of Line Calibration

Valeo (BYD) | 2021 - 2021
  • EOL Calibration using custom targets instead of checkerboards.
  • Finds known feature points and minimizes reprojection error.
  • Developed feature point detection and cost functions for custom targets.

Motion Tracking Calibration

Valeo (GM, FORD, BMW) | 2021
  • Dynamically estimates surround-view camera extrinsics while vehicle is moving.
  • Block-based motion tracking and cost function using motion vectors.
  • Maintained algorithm and implemented bug fixes.

Stereo Presentation Attack Detection

Vis-Ai Labs (Fujitsu) | 2019
  • Spoof detection system for ATM face authentication (Spain/Fujitsu).
  • Uses stereo camera depth to detect human face presence vs. spoofs (photos, tablets).
  • Prototyped VGG16 model using depth features, securing the project.

TaraXL - Stereo Camera

Vis-Ai Labs | 2018 - 2019
  • Outputs depth map along with image.
  • Uses pixel disparity and triangulation for depth.
  • Developed Semi-Global Matching (SGM) algorithm from scratch; optimized with CUDA for Jetson Nano.

EsomGears & Health Monitor

Vis-Ai Labs | 2016 - 2018
  • OTA and device stat monitoring for Linux/Android.
  • Uses Ostree (Linux) / native interfaces (Android) for OTA; MQTT to AWS IoT Hub for stats.
  • Developed cloud-side from scratch; worked on device monitoring daemon service.

Personal Project: Déjà Vu

2020 - 2021

An open-source cheatsheet project to document learnings in Computer Vision.

Params: 12.3M
Output: Project Showcase
Filter: Multi-Scale
Layer: Activation (ReLU)

Technical Skills

Core Technical

Computer Vision Deep Learning DNN Machine Learning Image Processing

Programming & Tools

C++ Python CUDA JavaScript OpenCV PCL PyTorch PyTorch-Lightning Keras TensorFlow Git Linux

General

Algorithm Design System Design Problem Solving Self Learning Team Player Innovative Curious Motivated

Awards & Recognition

Innovation Award Valeo (2022)
High Flyer Award (4) Valeo (2021-2023)
Employee Of the Month (6) Valeo (2021-2024)
Emerging Engineer Award E-con Systems (2019)
Params: 3.2M
Output: Skill Matrix
Filter: Academic
Layer: Normalization (BatchNorm)

Education

M.Tech Artificial Intelligence & Machine Learning (Deep Learning Specialization) 8.2 CGPA
BITS Pilani, WILP, India 2022 - 2025
BE Computer Science & Engineering 7.6 CGPA
PERI IT, Chennai, India 2012 - 2016
HSC Computer Science 88%
Jaigopal Garodia National, Chennai, India 2010 - 2012
Params: 1.5M
Output: Degree Stack
Filter: Social
Layer: FullyConnected (1024)

Get In Touch

Feel free to reach out via email or connect with me on social platforms.

Params: 4.8M
Output: Network Connect