<E.Justice/>
ML Researcher | LLM Inference Optimization | vLLM
Hi, I'm Ethan Justice - a Computer Science student at the University of Michigan researching KV-cache optimization and multi-agent LLM systems. Building production-grade ML infrastructure on HPC clusters.
02. Experience Timeline
From ML research to enterprise software engineering, building production systems at scale.
Machine Learning Researcher
Transitioning state-of-the-art KV-caching research into production-compatible infrastructure for vLLM and LMCache ecosystem.
AI Engineering Intern
Second non-founding engineer building core infrastructure for an agentic memory platform.
Software Engineering Intern
Developed and improved critical internal systems for operational efficiency across 5,000+ franchise locations.
Research Assistant
Created embedded systems for real-time user activity identification using machine learning.
Flight Systems Developer
Developed software for autonomous plane navigation and control surfaces.
Software Development Intern
Re-engineered internal employee rewards platform used by 750+ team leads.
Lead Programmer
Led software development for competitive FIRST Robotics Competition team, programming autonomous and teleoperated robot control systems.
03. Work & Projects
From ML research to hackathon wins, building solutions across the stack.
Efficient Heterogeneous LLM Multi-Agent Debate Framework
Resource-efficient Multi-Agent System reducing inference costs for complex reasoning tasks by 40%.
- ▹HPC Deployment on Great Lakes cluster with Slurm job scheduling
- ▹40% reduction in total FLOPs via confidence-based gating mechanism
- ▹Factory Pattern for modular model backend switching between vLLM and HuggingFace
- ▹Identified Syntactic Determinism failure modes in token-confidence calibration
Information Digest
Agentic information retrieval pipeline with Infrastructure-as-Code and CI/CD automation.
- ▹Automated infrastructure lifecycle using Terraform and GitHub Actions
- ▹Comprehensive Pytest suites before deploying to GCP Cloud Run
- ▹Real-time search pipeline with LangChain and SerpAPI
- ▹LLM agent autonomously decomposes queries and synthesizes answers
- ▹Strongly-typed REST interface with deterministic JSON outputs
Machine Learning Researcher
Transitioning state-of-the-art KV-caching research into production-compatible infrastructure for vLLM and LMCache ecosystem.
- ▹Productionizing research from 'Compute Or Load KV Cache? Why Not Both?' paper
- ▹Porting logic from legacy codebase to current vLLM and LMCache main branches
- ▹Enabling bidirectional KV prefill to reduce Time-to-First-Token (TTFT)
- ▹Optimizing GPU memory transfer overhead on Great Lakes HPC (A100/H100 clusters)
AI Engineering Intern
Second non-founding engineer building core infrastructure for an agentic memory platform.
- ▹Architected real-time Multimodal RAG pipeline using GCS and Redis
- ▹Enabled sub-second voice-to-voice agent interactions
- ▹Engineered semantic caching layer via Convex - 99% reduction in API response times
- ▹Built 'LLM-as-a-Judge' testing suite for agent memory evaluation
Software Engineering Intern
Developed and improved critical internal systems for operational efficiency across 5,000+ franchise locations.
- ▹Architected C# .NET microservice with REST API for SMS automation via Twilio
- ▹Single-handedly implemented complete CI/CD pipeline in Azure DevOps
- ▹Created batch-processing API achieving 90% reduction in developer time
- ▹Built franchise store hours compliance tracker for company-wide initiative
Interview Bot Pro
MacOS application for AI-powered interview practice with speech prosody analysis.
- ▹Tailored questions based on user experiences and target job requirements
- ▹Dynamic follow-up questions based on previous answers
- ▹STAR method structure analysis and content grading
- ▹Voice prosody analysis using Random Forest ML trained on MIT interview dataset
- ▹Job fit analysis with skill parsing and requirement matching
Ribbet - MHacks Project
Cross-platform social media app with betting mechanics built during MHacks hackathon.
- ▹MERN Stack architecture for cross-platform mobile development
- ▹Complete CRUD functionality for user and post data
- ▹Social media features: profiles, feeds, and betting mechanisms
The Situation Room - Cal Hacks
Unity game simulating de-escalation scenarios with AI-driven dialogue analysis.
- ▹Interactive dialogue with dynamic character responses
- ▹Google Gemini integration for dialogue content analysis
- ▹Hume AI for real-time sentiment analysis
- ▹Python backend server for API integration
Research Assistant
Created embedded systems for real-time user activity identification using machine learning.
- ▹Built embedded system on Orange Pi for activity level identification
- ▹Leveraged Llama 3 model for sensor data analysis
- ▹Implemented multithreading for performance optimization
- ▹Strategic task routing to high-performance CPU cores
Inner Voice AI
Conversational AI therapist that won 1st place at Stemist Hacks hackathon.
- ▹1st Place Winner at Stemist Hacks hackathon
- ▹OpenAI API for natural conversational flow
- ▹Hume AI for real-time sentiment analysis
- ▹Firebase for user sessions and conversation history
Software Development Intern
Re-engineered internal employee rewards platform used by 750+ team leads.
- ▹20% increase in monthly active users after platform modernization
- ▹Led requirement gathering with Employee Engagement stakeholders
- ▹Managed backend operations for 50,000+ entry dataset
- ▹Implemented automated testing protocols for high availability
Thread Library
User-level threading library with context switching
- ▹Context switching
- ▹Thread scheduling
- ▹Synchronization primitives
Pager
Virtual memory pager with demand paging
- ▹Page fault handling
- ▹Memory management
- ▹Swap space
File System
Multi-threaded network file server
- ▹File system design
- ▹Network protocols
- ▹Concurrency
Multi-Agent Debate Framework
Efficient heterogeneous LLM multi-agent debate for factual reasoning
- ▹40% FLOP reduction
- ▹Confidence-based gating
- ▹HPC deployment
Pipeline Simulator
5-stage MIPS pipeline simulator with hazard handling
- ▹Data hazards
- ▹Control hazards
- ▹Forwarding
Maze Solver
Puzzle solver using BFS/DFS with custom data structures
- ▹Search algorithms
- ▹Stack/Queue implementation
Priority Queue
Binary and pairing heap implementations
- ▹Heap operations
- ▹Amortized analysis
Database Query Optimizer
SQL-like query processor with indexing
- ▹B-tree indexing
- ▹Query optimization
TSP Solver
Traveling Salesman Problem with MST heuristics
- ▹MST algorithms
- ▹Branch and bound
- ▹Approximation
Flight Systems Developer
Developed software for autonomous plane navigation and control surfaces.
- ▹Implemented 3DVFH* algorithm for obstacle avoidance
- ▹Used ROS and MAVLink for plane control
- ▹Integrated remote ID for aircraft location tracking
MAT-5 Autonomous Plane
Flight systems software for autonomous plane navigation using 3DVFH* obstacle avoidance.
- ▹3DVFH* algorithm for dynamic obstacle avoidance
- ▹Autonomous path following
- ▹Remote ID for receiving locations of other aircraft
- ▹ROS and MAVLink for control surface integration
Reaper - FRC 2023
Competition robot with differential swerve drive for FRC Charged Up game.
- ▹Motion profiling for precise velocity and position control
- ▹Differential swerve drive chassis control
- ▹PID tuning with data collection and analysis
- ▹Autonomous pathing with time axis integration
Blade - FRC 2022
Swerve drive robot for FRC Rapid React with autonomous shooting up to 30ft.
- ▹Autonomous shooting with calculated launch angle and velocity
- ▹State-based control for robot operations
- ▹Time-of-flight sensor integration with Arduino
- ▹Absolute arm angle calculation using gravity vector
NavPod
Embedded system for precise robot position tracking using sensor fusion.
- ▹Optical flow sensor for translational movement tracking
- ▹IMU for rotational movement tracking
- ▹Custom sensor fusion algorithm for position determination
- ▹I2C communication with serial output over USB
Raptor - FRC 2020
Competition robot with autonomous turret tracking and Hermite spline pathing.
- ▹Hermite spline autonomous pathing
- ▹NavPod integration for precise localization
- ▹Auto-aiming turret based on robot position
- ▹Computer vision with GRIP pipeline for ball detection
Lead Programmer
Led software development for competitive FIRST Robotics Competition team, programming autonomous and teleoperated robot control systems.
- ▹Developed autonomous routines using path planning and odometry
- ▹Implemented vision processing for game piece detection
- ▹Programmed drivetrain, arm, and intake subsystems in Java
- ▹Mentored younger students in programming and engineering practices
04. Education
Coursework, projects, and academic achievements at Michigan Engineering.
University of Michigan
Bachelor of Science in Engineering in Computer Science
Aug 2023 - May 2026• Ann Arbor, Michigan