Hi, I'm Adithya Balaji
AI/ML Systems Engineer
I engineer resilient AI systems, agentic architectures with the Model Context Protocol (MCP), and Vercel AI SDK. My work spans hybrid RAG retrieval, latency-optimized LLM pipelines, and learned database optimizers with Tree-CNNs.
Building MCP tool servers and studying agentic tool-calling under an NSA-sponsored RF capstone.
Cut latency by 50% & token usage by 2x using Vercel AI SDK, guardrails, and deterministic routing.
Tree-CNN query optimization reducing PostgreSQL Join Order Benchmark latency by 25%.
About Me
I am a graduate student pursuing an MS in Computer Science at The University of Texas at Dallas. My work centers on building robust AI architectures that move beyond toy demos into high-reliability, deterministic production systems.
For my graduate capstone (CS6301), I am working on an NSA-sponsored INSuRE project exploring agentic emanations analysis. Using the Model Context Protocol (MCP), we build tool primitives for digital signal processing (DSP) to analyze how LLM agentic tool-calling behaves under complex, blind signal environments.
Prior to my graduate studies, I graduated with a B.Tech in CSE from the National Institute of Technology Allahabad and worked as a software consultant at PwC, engineering high-throughput automation pipelines, ETL layers, and CRM enterprise systems.
The University of Texas at Dallas
Master of Science in Computer Science
RF Capture-the-Flag Testbed for Agentic Emanations Analysis via Model Context Protocol.
National Institute of Technology Allahabad
B.Tech in Computer Science & Engineering
Career & Industry Roles
AI/ML Engineering Intern
Early-stage EdTech Startup (Stealth)
- Designed and implemented a "Struggle Signature" behavioral classification system using confidence scores, response latency, and a frustration index, combined with zero-temperature prompting for deterministic classification.
- Re-architected system prompt and orchestration logic, cutting end-to-end response latency by 50% and reducing token consumption per query by 2x.
- Built the system’s LLM guardrails from scratch, including reply gating and safety checks to prevent direct answer leakage, later extended by the lead architect.
- Built a CRUD API linked directly to Supabase for managing the curriculum topic graph, with batch upload via CSV/Excel, cycle detection, and duplicate checking to guarantee graph integrity.
Consultant / Software Developer
PricewaterhouseCoopers Services LLP
- Engineered C# and JavaScript automation plugins against Microsoft Dataverse using FetchXML queries, eliminating 15+ hours of manual processing weekly across enterprise CRM workflows.
- Built scheduled automation systems and multi-stage approval workflows using Power Automate, reducing turnaround time on client-facing processes by 30%.
- Built an Azure Pipeline to automate data extraction from SAP HANA into a Power Apps-based ETL layer, feeding Power BI dashboards for real-time KRI monitoring and operational analytics.
Course Grader / Student Assistant
The University of Texas at Dallas
- Grading assignments and hosting 4 weekly office hours for CS4349 Advanced Algorithm Design and Analysis, mentoring undergraduate and graduate students in complex algorithm analysis.
Featured Projects
Deep dives into learned systems, Model Context Protocol tooling, and high-throughput LLM architectures.
RAG-Powered AI Tutoring Platform
100% Top-2 Hit RateAgentic & RAG Architecture
Hybrid Retrieval & Latency-Optimized Architecture
- Built a hybrid retrieval engine over semantically-chunked curriculum content using all-MiniLM-L6-v2 embeddings.
- Fused dense vector similarity (Chroma) and sparse lexical search (BM25) using Reciprocal Rank Fusion (RRF).
- Orchestrated multi-step conversational guidance with LangGraph while maintaining sub-2-second response latency.
Bao: Learned Query Optimizer
25% Latency ReductionLearned Systems
Deep Learning for Relational Database Engines
- Re-implemented the Bao optimizer end-to-end in PyTorch to automate cost-model decisions in relational databases.
- Trained a Tree-structured Convolutional Neural Network (Tree-CNN) to predict plan performance directly from PostgreSQL EXPLAIN execution trees.
- Employed contextual bandits (Thompson Sampling) to balance exploration of alternative query plans with minimal runtime overhead.
INSuRE RF Capture-the-Flag
Academic CapstoneGraduate Capstone (CS6301) · NSA-Sponsored
Agentic Emanations Analysis via Model Context Protocol
- Architected a specialized Model Context Protocol (MCP) server that exposes digital signal processing (DSP) tool primitives to an autonomous LLM agent.
- Constructed a Capture-the-Flag (CTF) testbed simulating physical signal emanations to stress-test LLM planning in unconstrained RF environments.
- Studied how structured tool definitions and feedback loops influence an agent’s accuracy in blind signal classification.
Skills & Technologies
A comprehensive toolkit spanning agentic tool protocols, low-latency LLM orchestration, and distributed data systems.
Languages
Core languages for systems, APIs, and ML modeling
AI & Agentic Frameworks
Orchestration, tool calling protocols, and modern retrieval pipelines
ML, Deep Learning & Evaluation
Model training, learned data structures, and rigorous LLM evaluation
Data & Infrastructure
Production vector storage, relational engines, and cloud pipelines
Let's build something impactful.
I am actively seeking AI/ML engineering roles, agentic systems research collaborations, and ambitious technical challenges. Feel free to reach out directly.