Roshan V. Sathyanathan
[Address]
Education
Indian Institute of Information Technology, Design and Manufacturing, Kancheepuram
2021—Present
B. Tech Computer Science
Baldwin Boys High School, Bangalore
2009—2019
ICSE
Work Experience
Vectrum Technologies
2023—Present
Technical Lead
- Engineered a semantic search system for a large document corpus, able to effectively retrieve relevant extracts.
- Re-engineered existing server-based application backend with a serverless architecture, enabling automatic global scaling and reducing costs and operational overhead.
- Optimized dynamic routing and streaming from multiple LLM endpoints, cutting end-user latency, improving throughput and mitigating downtime under peak loads.
- Fine-tuned BART with a classification head to serve as a zero-shot text classifier, and deployed the model as a containerized microservice for inference in production.
Vectrum Technologies
2022—2023
Full Stack Engineer
- Designed and developed several applications for the web with NextJS.
- Architected and built a microservices-based backend with Express on NodeJS.
- Deployed and maintained applications and services (MongoDB, Redis) on AWS, while optimizing for cost.
- Set up continuous integration pipelines for automated testing and deployment, with Github Actions and Jenkins.
AUV Society, IIITDM
2022—2024
CS Subsystem Lead
- Led development of the autonomy stack that powered the team’s top-40 finish at SAUVC 2023 and a top-10 placement in 2024, including simulation, perception, and control components.
- Set up a ROS-compatible underwater simulator/testbench that mirrored real-world dynamics, enabling end-to-end testing of navigation and control systems.
- Implemented rudimentary underwater SLAM using stereo vision and a depth sensor under severe visibility constraints, along with classical CV techniques (HOG, edge and line detection) for object recognition.
- Architected and implemented low-level motor control with compensation for non-linear thrust characteristics across several heterogeneous motors, and along with a high-level DoF based control system with an integrated PID controller for stable underwater movement.
Projects
LLMs for Document Information Extraction from Visually Rich Documents
Final Year Project
- Improved on the LMDX paper by proposing a better prompting strategy, paired with a constrained decoding approach to reduce hallucinations and inaccuracies.
- Achieved a slight improvement in performance over the current best achieved by the original paper.
Deep RL for 3D Bipedal Robot Locomotion in MuJoCo
Reinforcement Learning Course Project
- Implemented PPO, A2C, and SAC algorithms to train a bipedal robot for stable locomotion
calc9000: A makeshift implementation of Mathematica/WolframScript
Hobby project
- Designed an incomplete parser that can generate an AST from valid Wolfram Language code.
- Built an interpreter that provides a REPL environment, and can run also scripts.
- Implemented several built-in operators and functions, and provide basic computer algebra capabilities with NumPy/SymPy.
Coursework
Core Courses
Operating Systems
Computer Networks
Compiler Design
Computer Organization and Architecture
Database Systems
Object-Oriented Programming
Digital System Design
Design and Analysis of Algorithms
Basic Math
Calculus
Differential Equations
Linear Algebra
Probability and Statistics
Design and Management
Sociology of Design
Systems Thinking for Design
Smart Product Design
Entrepreneurship and Management Functions
Prototyping and Testing
Specialized Electives
Data Science
Introduction to Machine Learning
Essential Mathematics for Machine Learning
Deep Learning
Reinforcement Learning
Digital Image Processing
Introduction to Biometrics
IIoT and Cloud Computing
High-Performance Computing
Technical Skills
- Experienced with ROS and Robotics Systems
- Proficient in JavaScript, Python, Go and C/C++
- Familiar with PyTorch, TensorFlow, Python DS stack (NumPy, Pandas, Matplotlib), CUDA