About Me
// TL;DR
MSCS Grad '25 @ Iowa State. Full Stack Engineer & AI Researcher. Ship production apps with Next.js, Supabase & CI/CD. Build autonomous AI content pipelines. Steer LLM internals with PyTorch. Seeking SWE / Full Stack / AI roles.
I am Devendra Sai Mupparaju, a Master of Science in Computer Science graduate from Iowa State University (December 2025). I build full-stack platforms end-to-end, engineer autonomous AI systems, and research ways to make AI models more reliable — from production Next.js deployments and multi-API automation pipelines to training-free LLM steering techniques.
Full Stack Engineering
Architected Grocery Square, a production-grade e-commerce platform with Next.js App Router and Supabase/PostgreSQL, deployed via GitHub Actions CI/CD to GitHub Pages with SSG pre-rendering across thousands of product routes.
Built Hobby Hive, a real-time social platform using React and Supabase, featuring instant video updates through a Time Capsule architecture and dynamic TypeScript dashboards for data visualization.
Engineered Bounce Streak during a hackathon — a 2D arcade game on Reddit Devvit serving 3,000+ concurrent users at 60 FPS, with deterministic daily challenges via seeded randomization in Phaser and TypeScript.
AI Engineering & Research
Developed a training-free Activation Steering method in PyTorch that injects a correctness vector into CodeLLM residual streams during inference, boosting CodeLlama-7B repair accuracy from 48% to 56% and eliminating invalid code generation entirely.
Secured deep learning training loops inside Intel SGX enclaves, isolating computation from OS-level vulnerabilities and maintaining 100% data privacy during simulated system compromise scenarios.
Formally verified fairness properties of a federated learning protocol using NuSMV and CTL, modeling the 3-client workflow as a finite-state machine to prove no client starvation or deadlocks occur across aggregation rounds.
Autonomous AI Systems
Built an end-to-end autonomous content pipeline ("The Social Speaker") that researches trending topics via Perplexity API, generates 3-slide Instagram carousel narratives with Gemini 2.5-Flash, creates 1080×1080 dark-mode visuals using Imagen 3, and auto-publishes to Instagram — all without manual intervention.
Engineered a multi-tier image generation fallback chain (Imagen 3 → Pollinations → Pillow) and integrated 4 external APIs with retry logic and structured CSV logging, scheduled daily via Windows Task Scheduler.
Technical Skills
Languages
LANG_MODULECloud & Ops
OPS_MODULEMachine Learning
AI_MODULEDatabases
DB_MODULEWeb & Tools
WEB_MODULEKey Concepts
CORE_MODULEProject Architecture
GROCERY SQUARE
Architected a production-grade full-stack grocery e-commerce platform deployed to GitHub Pages via automated CI/CD pipelines. Leveraged Next.js SSG with generateStaticParams to pre-render product routes at build time, eliminating runtime 404 errors at scale. Engineered a resilient fallback architecture toggling between live Supabase/PostgreSQL and local mock data for zero-downtime deployments across environments.
VIEW_PROJECT
HOBBY HIVE
Validated real-time data scalability by architecting a 'Time Capsule' feature with Supabase, ensuring instant video updates. Established a seamless user experience strategy by engineering dynamic dashboards in TypeScript to visualize complex hobby progress data intuitively.
VIEW_PROJECT
AI SOCIAL MEDIA GENERATOR
Autonomous content pipeline that pulls trending AI topics via Perplexity API, generates captions with Gemini 2.5-Flash, creates images using Imagen 3, and auto-publishes to Instagram — fully scheduled with zero manual steps. Features a multi-tier fallback chain for image generation (Imagen 3 → Pollinations → Pillow), a custom slide compositor with topographic texture, gradient typography, and rounded image cards, plus 4 external APIs integrated with retry logic and structured CSV logging.
VIEW_PROJECTLIVE_DEMOACTIVATION STEERING APR
A training-free method for enhancing CodeLLM reliability in Automatic Program Repair. Computes a correctness vector from contrastive buggy/fixed code pairs and injects it into the model's residual stream during inference. Eliminated invalid code generation entirely and boosted CodeLlama-7B accuracy from 48% to 56%.
VIEW_PROJECTFEDERATED FAIRNESS VERIFICATION
Formal verification of fairness properties in a 3-client federated learning protocol. Models the FL workflow as a finite-state machine in NuSMV, encodes CTL properties for liveness, safety, and fairness, then automatically verifies no client is starved and no deadlocks occur across bounded aggregation rounds.
VIEW_PROJECTPRIVACY PRESERVING
USING INTEL SGX
Proved the integrity of sensitive training loops in untrusted environments. By leveraging Intel SGX enclaves, this architecture isolates computation from OS-level vulnerabilities, ensuring privacy was maintained even during simulated system compromise scenarios.
BOUNCE STREAK (HACKATHON)
Engineered a high-performance 2D arcade physics engine optimized for mobile browsers. Implemented 'cushion' collision detection to deliver a smooth 60 FPS experience. Architected a deterministic daily challenge system using seeded randomization, ensuring identical physics conditions for thousands of concurrent users.
VIEW_PROJECT
Education
Iowa State University
Aug 2023 - Dec 2025Focus on Artificial Intelligence, Machine Learning, and System Architecture. GPA: 3.5/4.0
SRM University Amaravati
2019 - 2023Focus on Software Engineering, Data Structures, and Algorithms. GPA: 7.8/10
Initialize Connection
Ready to collaborate? Establish a secure link below.