Hi there, I’m

Sai Manvith

Hi there

My name is Sai Manvith Chatrathi. I'm a Computer Science student at UMass Boston, focused on artificial intelligence, machine learning, and software engineering. I'm particularly interested in reinforcement learning, language models, and intelligent systems that learn from interaction and feedback.

My interests center on how learning-based systems can reason, adapt, and make better decisions. Alongside AI/ML, my work spans software engineering, full-stack development, and embedded systems, allowing me to approach technical problems from both research and engineering perspectives.

My current interests include reinforcement learning and modern foundation models as I prepare for graduate research in artificial intelligence and machine learning. Below are selected projects that reflect my research interests and engineering experience.

Projects

Multi-Agent Research Platform

AI Agents · LLM Systems · Research Automation

I developed a full-stack multi-agent AI research platform that decomposes complex research workflows across specialized agents for planning, information retrieval, analysis, and synthesis. Rather than relying on a single language-model interaction, the system coordinates five or more specialized agents to execute multi-step research tasks.

I implemented agent orchestration, tool calling, contextual memory, semantic search, and retrieval-augmented generation using LangGraph and LangChain. FastAPI services connect the agent layer with application workflows, while vector search and PostgreSQL-backed pipelines support analysis across more than 25,000 indexed documents.

I designed the project to explore how language-model agents can decompose difficult problems, retrieve external evidence, maintain useful context, and coordinate decisions across longer workflows. This work directly shaped my broader research interests in agentic AI, foundation models, interactive learning, reinforcement learning, and feedback-driven adaptation.

AI Document Intelligence System

Document AI · Multimodal Learning

I developed an AI-powered document intelligence platform for automated document understanding, classification, and structured information extraction across more than 100,000 records.

I built OCR and transformer-based processing pipelines using LayoutLM and PyTorch, allowing the system to model both textual content and spatial document structure. The resulting pipeline achieved approximately 91% extraction accuracy on structured document fields.

I also implemented automated evaluation and processing workflows that reduced manual document-review effort by approximately 65%. The project strengthened my interest in multimodal representations and in systems that must reason over several forms of context simultaneously.

Diabetes Decoded

Applied Machine Learning · Full-Stack AI

I developed an end-to-end healthcare analytics platform for securely managing health information, identifying trends, and generating predictive insights from more than 1,000 user health records.

I built machine-learning analytics pipelines that achieved approximately 87% prediction accuracy in my evaluation and connected them to a full-stack application using Next.js, Flask, Firebase, and REST APIs. Interactive dashboards expose health trends and model outputs while authentication and persistent storage support secure profile management.

I also integrated visualization and external services for diet and exercise information. Building the system reinforced my view that successful machine learning depends not only on model performance but also on reliable data pipelines, software architecture, security, and understandable interfaces.

Real-Time Collaboration Platform

Distributed Systems · Real-Time Software

I engineered a collaborative workspace supporting real-time editing, comments, authentication, and team collaboration. Persistent WebSocket connections provide low-latency bidirectional communication, while Redis Pub/Sub distributes events across application instances.

I designed REST APIs, role-based access control, and PostgreSQL schemas for secure collaboration workflows. The real-time architecture was designed to support more than 1,000 concurrent connections while maintaining synchronized application state across connected clients.

This project gave me practical experience with distributed state, concurrency, event propagation, synchronization, and the engineering tradeoffs involved in building interactive systems at scale.

Distributed URL Shortener

Distributed Systems · Backend Engineering

I developed a distributed URL-shortening service with REST APIs for URL generation, redirection, and analytics tracking. The system was designed around scalable backend architecture rather than treating URL shortening as a simple database lookup.

I implemented Redis caching and rate limiting to reduce repeated database access and protect backend resources, reducing average response latency by approximately 60% in my testing. Persistent URL data and analytics are managed through PostgreSQL.

I containerized the services with Docker and deployed them on AWS using a scalable microservices-oriented architecture. The project allowed me to explore how caching, persistence, API design, rate limiting, and deployment decisions influence system behavior under increasing demand.

Real-Time Chat Platform

Networking · Distributed Backend Systems

I built a real-time messaging platform designed to support more than 1,000 concurrent WebSocket connections while maintaining responsive communication between connected users.

I implemented JWT-based authentication and Redis Pub/Sub for distributed message delivery, reducing message-delivery latency by approximately 50% in my testing. I also designed backend APIs, PostgreSQL schemas, session management, and user and conversation workflows.

The project gave me hands-on experience with persistent connections, asynchronous communication, distributed messaging, session state, and the challenges involved in maintaining low-latency behavior as concurrent usage grows.

VoteSecure — Secure Online Voting Backend

Backend Engineering · Authentication

I developed the backend for an online voting application using Node.js and Express.js, with an emphasis on authenticated access, structured voter and candidate data, and reliable API design.

I implemented JWT-based authentication and Aadhaar-style identity verification logic for access control and designed MongoDB schemas for voter and candidate management. I also deployed the application online with attention to data security and backend reliability.

The project gave me experience thinking about authentication, authorization, data modeling, and reliability in applications where access to sensitive operations must be carefully controlled.

Linux GPIO Character Device Driver

Operating Systems · Embedded Linux

I developed a Linux character device driver in C for direct Raspberry Pi GPIO control using the Linux Kernel Module framework, working directly across the boundary between user-space applications, kernel-space software, and physical hardware.

I implemented ioctl() interfaces for GPIO configuration and control, created user-space applications for interacting with the driver, and developed Makefiles for compiling, loading, and deploying the kernel module on Embedded Linux.

I validated driver functionality using Linux kernel logs through dmesg, system utilities, and command-line debugging tools. Building the driver gave me a deeper understanding of operating-system abstractions and the software interfaces connecting applications to hardware.

STM32 Environmental Monitoring System

Embedded Systems · Real-Time Firmware

I developed Embedded C firmware for an STM32-based environmental monitoring system that acquires and processes sensor measurements in real time.

I integrated temperature and humidity sensors through I²C and implemented ADC-based analog signal acquisition. Hardware timers and interrupts drive periodic sensor sampling and UART communication, avoiding dependence on a purely sequential polling architecture.

I also implemented low-power firmware techniques and validated system behavior using STM32CubeIDE debugging and serial communication tools. The project strengthened my understanding of timing, interrupts, resource constraints, and hardware-software coordination in real-time systems.

ESP32 Smart Home Controller

FreeRTOS · IoT · Concurrent Embedded Systems

I developed multitasking firmware using FreeRTOS for an ESP32-based smart-home automation system, separating sensing, communication, and device-control responsibilities into independently scheduled tasks.

I implemented MQTT communication over Wi-Fi for real-time cloud connectivity and remote monitoring and integrated environmental sensors and actuators through GPIO and SPI interfaces.

I designed inter-task communication using FreeRTOS queues, semaphores, and software timers and optimized memory utilization and task scheduling to improve responsiveness. The project provided practical experience with concurrency and synchronization under embedded resource constraints.

RTOS-Based Smart Traffic Controller

Real-Time Systems · Embedded Control

I developed a multitasking traffic-signal controller using STM32 and FreeRTOS to simulate a real-time traffic-management environment with several independently executing control requirements.

I implemented concurrent RTOS tasks for traffic lights, pedestrian crossings, and emergency-vehicle priority handling. GPIO, hardware timers, and external interrupts provide deterministic signal control, while FreeRTOS queues, semaphores, and software timers coordinate synchronization and communication between tasks.

I optimized scheduling latency and resource utilization to maintain reliable real-time behavior. The project gave me deeper experience reasoning about deterministic execution, task priorities, concurrency, synchronization, and resource management in embedded systems.

My projects span language-model agents, multimodal AI, machine learning, distributed systems, backend engineering, operating systems, and real-time embedded computing. Across these areas, I am especially interested in building intelligent systems that can reason over information, learn from interaction and feedback, make decisions under uncertainty, and remain reliable when deployed in complex real-world environments.

This represents just a small selection of the projects I have worked on over the years. I have many more pieces of personal, university and design work.

Experience

AI Research Assistant

University of Massachusetts Boston · Jan 2025 – Present

I work on machine-learning systems for analyzing EEG and EMG signals in neurorehabilitation research, with a focus on translating physiological sensor data into computational representations that can support motor-intent prediction. My work spans the complete research pipeline, from data preprocessing and feature extraction to model development, evaluation, inference, and visualization.

I developed deep-learning pipelines using PyTorch and TensorFlow for real-time analysis across more than 10,000 experimental data samples and built and evaluated models for motor-intent classification, achieving approximately 85% classification accuracy. I also optimized preprocessing, feature-extraction, and inference workflows, reducing processing time by approximately 30%.

Beyond model development, I built React.js and Flask-based visualization and monitoring tools that allow research teams to inspect model outputs and rehabilitation metrics. I have worked across machine learning, backend services, data pipelines, and user-facing research software, giving me experience connecting experimental models with usable end-to-end systems.

This research has shaped my interest in intelligent systems that must learn from noisy real-world observations, generalize beyond individual examples, and produce decisions that can be evaluated in meaningful application settings.

Software Engineering Apprentice — PACE Program

University of Massachusetts Boston · Aug 2024 – Present

I contribute to the development of a real-time 3D indoor and outdoor navigation platform that combines interactive software, routing algorithms, wireless localization, and distributed embedded devices. The system connects React.js and Three.js visualization with Python-based services and physical sensing infrastructure.

I developed core navigation features and routing services for real-time path computation and integrated BLE, RFID, and Raspberry Pi mesh nodes for localization. This integration improved location accuracy by approximately 25% while requiring coordinated development across frontend software, backend APIs, networking, and embedded hardware.

I also built and maintained backend services and REST APIs and worked with Dockerized microservices, NGINX, and Firebase-based deployment workflows. These improvements reduced deployment overhead by approximately 35% while supporting greater system availability and reliability.

Working on PACE has given me experience with intelligent systems that operate beyond a purely digital environment: software must interpret observations from physical infrastructure, compute useful actions in real time, and remain robust as sensing and environmental conditions change.

AI Research Intern — Stanford University Coursework

Stanford Datasets · May 2024 – Jul 2024 · Remote

Through Stanford University coursework using Stanford datasets, I worked with large-scale machine-learning and language-model pipelines using PyTorch, TensorFlow, and Hugging Face. My work focused on model experimentation, optimization, evaluation, and understanding how training choices affect model behavior across datasets.

I fine-tuned language models and implemented parameter-efficient and feedback-based optimization techniques including LoRA and RLHF, reducing model training time by approximately 40%. I also developed automated evaluation, benchmarking, and monitoring workflows for comparing model performance across experiments.

I built tools for model evaluation and bias analysis and examined robustness, reliability, and generalization across multiple datasets. This work gave me practical exposure to the challenges involved in adapting foundation models efficiently while evaluating more than aggregate model performance alone.

The experience contributed directly to my current research interests in reinforcement learning, learning from feedback, foundation-model behavior, and the design of AI systems that can adapt their behavior through interaction and evaluation.

Software Engineering Intern

North South GIS India Pvt. Ltd · Aug 2023 – Dec 2023 · Remote

I developed backend services and RESTful APIs for a search and geospatial query platform using Python, JavaScript, PostgreSQL, and geospatial data services. My work focused on improving the systems responsible for retrieving, ranking, and serving location-oriented information.

I designed and optimized search and ranking functionality, improving search relevance and reducing failed queries by approximately 20%. I used A/B testing and performance analysis to evaluate changes empirically rather than relying only on implementation-level assumptions.

I also optimized backend workflows using asynchronous processing, improving service reliability and responsiveness by approximately 30%. The experience strengthened my foundation in backend engineering, API design, experimentation, performance optimization, and building software around real-world data.

Across my research and engineering experience, I have worked at multiple layers of intelligent systems: from physiological data and machine-learning models to language-model optimization, real-time navigation, distributed software, APIs, and embedded devices. These experiences have shaped my interest in building AI systems that can learn from data and feedback, reason under uncertainty, interact with complex environments, and remain reliable when deployed in real-world settings.

Contact

I'm always interested in discussing research, graduate opportunities, AI/ML, software engineering, and challenging technical problems. I'm particularly interested in opportunities to contribute to meaningful research and build intelligent systems with real-world impact. Feel free to reach out.

Reach out on Email, LinkedIn, or Phone.

Booting...

Starting...