Krishna Prashanth Thummanapelly

Supporting students in building cutting-edge AI projects in Health Care domain under Prof. Gondy Leroy.

- Architected and deployed an NLP-driven diagnostic web platform for Autism Spectrum Disorder detection from Electronic Health Records (EHR), processing structured and unstructured clinical data. - Optimized Python inference pipeline by eliminating redundant preprocessing and vectorization steps, improving API response latency by 10x (seconds → sub-second). - Containerized the application using Docker and deployed on AWS EC2 and University of Arizona High Performance Computing infrastructure, enabling scalable, reproducible research-grade infrastructure increasing throughput from 5 rpm to 30 rpm. - Designed and implemented a medical text simplification service using Spring Boot and transformer-based LLMs to reduce clinical reading complexity for non-expert users. - Integrated OpenAI-based LLM backend with prompt-engineered persona controls, enabling customizable simplification for caregivers, clinicians, and general audiences. - Reduced average text reading ease level by ~30-40% while preserving clinical meaning. - Developing AI-driven “digital twin”(avatars) conversational agents to simulate therapeutic interaction scenarios for autism support applications. - Building Agentic AI system for evaluation of mental health information queries in an automated fashion, improving query response coherence and contextual grounding using tool-augmented LLM reasoning.

- Teaching Assistant for the course - Introduction to Parallel and Distributed Programming by Prof. David Lowenthal. - Built automated grading framework for 70+ students, executing 500+ multithreaded and OpenMPI test cases per assignment cycle. - Implemented deterministic race-condition detection tests, correctness and performance tests reducing manual grading time by ~80%. - Cut assignment evaluation turnaround time from 5 days to 24 hours.

- Deployed autoscaling open-source LLM inference services on multi-GPU AWS EC2 (4–8 GPUs/node) using Ray Serve + FastAPI, supporting 1K+ concurrent requests/day. - Increased GPU utilization by ~35% via request batching and dynamic worker scaling, reducing p95 inference latency by ~40%. - Integrated fine-tuned Llama 3 into production workflows, eliminating third-party API dependency and reducing inference costs by ~60-70%. - Provisioned secure self-hosted Git server on EC2 with role-based access control, improving internal code security and reducing external repository reliance.

- TEFAR: An Efficient Transparent Finer-grained Encryption of Internet Access Artifacts - Presented at IEEE COMSNETS IIGW, Jan 2024, received *Best Paper Award* - Designed fine-grained encryption framework for large-scale internet log data storage with <0.1% encryption overhead. - Reduced encryption overhead compared to baseline approaches while preserving forensic-readiness compliance. - Validated performance across scalable big-data environments and columnar file formats like ORC.

- Achieved 30-minute update intervals with a low 50ms query latency by building a real-time large-scale app-based business analytics data querying system using Apache Pinot. - Utilized PySpark ETL for data transformation, aggregation, and ingestion. - Collaborated with Frontend developers, Data platform team and Data Analysts to design, develop, test and deploy the complex system and accomplish the project goals. - Enhanced system reliability by erasing discrepancies in delivery partner joining and referral bonus processes, resulting in 100% reduction of data mismatches between microservices. - Improved fraud detection with a real-time ML model that processed 16,000 rpm at 500ms latency, leveraging Amazon SQS and Shoryuken for efficient asynchronous data handling, while maintaining minimal lag. - Implemented real-time monitoring with updates every 100ms by integrating Prometheus metrics through a statsd sidecar or a microservice written in go and configuring Grafana for service sanity tracking. - Reduced app request timeouts from 10% to 1% by parallelizing third-party API calls using goroutines. - Increased system reliability by migrating Online Kafka usage across multiple microservices to Confluent’s Kafka cluster with 100% service uptime. - Communicated with multiple teams across the company to ensure the safe Kafka migration in services. Maintained documentation for best practices, testing, and deployment strategy for Kafka migration.

- Integrated Google reCAPTCHA v3, reducing spam registration attempts by ~70% in the website using PHP, JavaScript, JQuery - Developed Android-based patient intake application using Flutter, digitizing manual workflows and reducing manual entry errors. - Enabled structured export of 5K+ patient records into organized digital format. - Coordinated and adapted to changing business goals to meet the customer requirements.