<E.Justice/>

Optimization of Large Language Model inference infrastructure focusing on KV-cache management.

• ▹Productionizing research from 'Compute Or Load KV Cache? Why Not Both?' paper into the vLLM ecosystem
• ▹Implementing bidirectional KV prefill logic to minimize Time-to-First-Token (TTFT) latency
• ▹Porting experimental logic from legacy codebases to current main branches of vLLM and LMCache
• ▹Optimizing GPU memory transfer overheads on Great Lakes HPC (A100/H100) clusters

Core infrastructure engineer for an agentic memory platform; built the complete multimodal interaction pipeline.

• ▹Engineered the entire end-to-end multimodal pipeline, enabling seamless file uploads and voice interfaces
• ▹Achieved sub-second voice-to-voice agent interactions through optimized audio processing streams
• ▹Implemented a semantic caching layer via Convex, reducing API response times by over 99%
• ▹Built an 'LLM-as-a-Judge' testing suite to quantitatively evaluate agent memory performance

Research framework and class project reducing inference costs for complex reasoning tasks by 40% via heterogeneous agents.

• ▹Achieved 40% reduction in total FLOPs using a confidence-based gating mechanism for agent responses
• ▹Deployed high-performance inference pipelines on Great Lakes HPC cluster using Slurm scheduling
• ▹Implemented a Factory Pattern to modularly switch backends between vLLM and HuggingFace
• ▹Identified 'Syntactic Determinism' failure modes in token-confidence calibration

Agentic information retrieval pipeline with robust Infrastructure-as-Code and CI/CD automation.

• ▹Automated complete infrastructure lifecycle using Terraform and GitHub Actions
• ▹Engineered an LLM agent that autonomously decomposes complex queries and synthesizes answers
• ▹Developed comprehensive Pytest suites ensuring reliability before deployment to GCP Cloud Run
• ▹Designed a strongly-typed REST interface guaranteeing deterministic JSON outputs

High-performance microarchitectural simulator modeling a 5-stage pipelined CPU with a configurable cache hierarchy.

• ▹Engineered a 5-stage pipeline simulator (IF, ID, EX, MEM, WB) achieving cycle-accurate execution of LC-2K binaries
• ▹Implemented hazard detection logic to handle data dependencies via forwarding units and stall injection
• ▹Integrated a unified instruction/data cache with configurable associativity, block size, and write-back/allocate-on-write policies
• ▹Optimized memory access simulation using Least Recently Used (LRU) eviction to minimize miss rates

End-to-end translation system converting assembly code into executable machine binaries via intermediate object files.

• ▹Developed a two-pass assembler generating object files with symbol tables and relocation entries for global label resolution
• ▹Built a static linker capable of combining multiple object files, resolving cross-file dependencies, and managing stack allocation
• ▹Wrote optimized LC-2K assembly algorithms, including a recursive combination function managing stack frames and return addresses
• ▹Implemented bitwise simulation of the LC-2K ISA to validate machine code execution behavior

Backend development and DevOps automation for enterprise Computer Systems serving 5,000+ franchise locations.

• ▹Architected a C# .NET microservice to handle SMS automation at the scale of 5,000+ stores
• ▹Designed and implemented full CI/CD pipelines in Azure DevOps to automate deployment workflows
• ▹Created a batch-processing API that achieved a 90% reduction in developer time for data management tasks
• ▹Built compliance tracking tools for franchise store hours using MongoDB and Docker

Facial recognition system built using ResNet-18 and non-parametric instance discrimination.

• ▹Fine-tuned a ResNet-18 backbone using Contrastive Loss to optimize feature embeddings for facial identity
• ▹Implemented Non-Parametric Instance Discrimination (NPID) to maintain a memory bank of feature vectors
• ▹Visualized high-dimensional embedding clusters using t-SNE to verify class separation
• ▹Engineered a custom CNN architecture to benchmark against pre-trained models on a held-out dataset

AI-powered interview practice application featuring auditory analysis and content grading.

• ▹Trained a Random Forest ML model on the MIT interview dataset for voice prosody analysis
• ▹Integrated Google Gemini to grade answers based on STAR method structure and job requirements
• ▹Developed logic for dynamic follow-up questions tailored to previous user responses

Seamless image combination engine using Gaussian and Laplacian pyramids.

• ▹Constructed 4-level Laplacian Pyramids to decompose images into distinct frequency bands
• ▹Implemented seamless blending masks to combine images without visible seams or artifacts
• ▹Reconstructed high-fidelity final images by collapsing pyramid levels upsampled via Gaussian kernels

Advanced edge detection system using gradient filters and convolution math.

• ▹Designed steerable filters to detect edges at arbitrary angles (pi/4, pi/2, 3pi/4)
• ▹Implemented Gaussian and Box filters to analyze noise reduction trade-offs in edge detection
• ▹Calculated gradient magnitude and orientation maps to visualize structural image features

Concurrent network file server using fine-grained locking and crash-consistent disk logging.

• ▹Implemented fine-grained reader/writer locks (`boost::shared_mutex`) for high-concurrency file access
• ▹Ensured file system consistency and crash recovery via strictly ordered disk writes (inode vs. data)
• ▹Built hierarchical directory management and inode allocation logic
• ▹Designed thread-safe network communication using TCP sockets

External pager managing virtual address spaces with eviction policies and copy-on-write sharing.

• ▹Implemented the 'Clock' page replacement algorithm to approximate LRU eviction efficiency
• ▹Engineered copy-on-write optimizations for `fork` operations to minimize physical memory usage
• ▹Managed swap-backed vs. file-backed pages and eager swap reservation
• ▹Handled page faults and memory protection bits to simulate hardware MMU behavior

Unity-based simulation game using AI for dialogue analysis in de-escalation scenarios.

• ▹Integrated Hume AI for real-time sentiment analysis of user voice input
• ▹Utilized Google Gemini to dynamically generate NPC responses based on conversation history

Cross-platform social media application with betting mechanics built during MHacks.

• ▹Architected MERN Stack backend to support cross-platform mobile functionality
• ▹Implemented full CRUD operations for user profiles, feeds, and betting transactions

Embedded ML development for real-time user activity sensing and identification.

• ▹Engineered an embedded system on Orange Pi for real-time activity level identification
• ▹Optimized Llama 3 inference on edge hardware using strategic core routing and multithreading
• ▹Processed high-frequency sensor data streams for immediate classification

High-performance threading library supporting context switching and preemptive scheduling.

• ▹Implemented thread initialization and context switching using `getcontext`, `makecontext`, and `swapcontext`
• ▹Built synchronization primitives including Mutexes and Condition Variables with wait queues
• ▹Handled timer interrupts for preemptive thread scheduling and CPU time slicing
• ▹Designed RAII wrappers for automatic resource management and deadlock prevention

Conversational AI therapist focusing on mental health, winner of 1st place at Stemist Hacks.

• ▹Won 1st Place at Stemist Hacks for best overall project
• ▹Engineered natural conversational flow using OpenAI API and Hume AI emotional analysis
• ▹Implemented session management and history tracking using Firebase

Full-stack modernization of internal employee rewards platforms used by 750+ team leads.

• ▹ drove a 20% increase in monthly active users by re-engineering the legacy rewards platform
• ▹Managed backend operations and Apex controllers for a dataset exceeding 50,000 entries
• ▹Implemented automated testing protocols ensuring high availability for critical internal tools

Autonomous navigation software development for M-Fly Aero Design competition planes.

• ▹Implemented 3DVFH* algorithm for dynamic aerial obstacle avoidance
• ▹Developed autonomous control loops using ROS and MAVLink for actuator management
• ▹Integrated remote identification Computer Systems for real-time aircraft tracking

Technical lead for FRC Team 3536; architected software for 4 competitive robots (Reaper, Blade, Raptor, NavPod).

• ▹Designed 'NavPod' custom sensor fusion system combining Optical Flow and IMU data for precise localization
• ▹Implemented Hermite spline pathing for complex autonomous trajectory generation
• ▹Programmed advanced drivetrain logic including Differential Swerve and field-oriented control
• ▹Developed autonomous turret tracking using computer vision and physics-based launch calculations
• ▹Mentored student engineers in C++, object-oriented design, and control theory