JOHN GIMBA.Resume

Computer Science at Purdue · May 2027

I build reliable AI systems and products that make complex workflows feel clear.

I’m John, a software engineer working across agentic workflows, inference infrastructure, distributed systems, and customer-facing tools.

Explore my work Get in touch
agent.run / production healthy
01
Plan task3 tools selected
42 ms
02
Retrieve contexthybrid search · 8 chunks
118 ms
03
Execute tools2 parallel · 1 approval
640 ms
04
Evaluate responsegrounded · passed
91 ms
checkpoint savedtrace replay available
2×AWS internships
4featured systems projects
May ’27Purdue CS graduation
U.S.citizen

01 / Selected work

Systems designed for real constraints.

Each project starts with a concrete bottleneck, then connects the implementation to measurable behavior.

012026

AI agents · Full stack

Agentic Workflow Orchestrator

A fault-tolerant runtime for multi-step agents with typed tools, human approvals, and replayable execution traces.

  • Engineered an asynchronous Python state machine that coordinates parallel tool calls and persists workflow state.
  • Added retries, timeouts, idempotency, checkpoint recovery, and policy controls for dependable execution.
  • Created a React and TypeScript console backed by GraphQL to inspect runs, approve actions, and debug failures.
PythonReactTypeScriptGraphQL
022026

RAG · Evaluation

Agent Memory and Evaluation Suite

A retrieval and evaluation system that gives long-running agents useful context without replaying complete histories.

  • Combined vector and keyword retrieval with metadata filters and checkpointed conversation state.
  • Measured retrieval precision, context recall, groundedness, and answer consistency with replayable evaluations.
  • Used cache invalidation and run replay to detect regressions after prompt or model changes.
PythonRAGVector SearchEvaluation
032025

LLM systems · Performance

LLM Inference Gateway

A model-serving gateway built to keep interactive latency stable as traffic and backend conditions change.

  • Implemented streaming, continuous batching, caching, rate limits, model routing, and circuit breakers.
  • Benchmarked Llama 3 70B with vLLM and TensorRT-LLM across four Purdue H100 GPUs.
  • Scaled per-user throughput from 22 to 118 tokens per second while reducing p99 time to first token to 195 ms.
vLLMTensorRT-LLMH100Load Testing
042026

CUDA · GPU optimization

Fused Attention Kernel

A custom FlashAttention-2 kernel that reduces memory traffic during long-context inference on an A100 GPU.

  • Fused attention operations in Triton and CUDA for Llama 3 8B at an 8,192-token context.
  • Reduced attention latency from 2.65 ms to 1.15 ms per step compared with eager PyTorch.
  • Avoided the intermediate attention matrix and reduced peak VRAM use from 14.2 GB to 2.1 GB.
CUDATritonPyTorchA100

AWS · Production systems · 2026

Certificate Authority Pool Management

Customers needed to rotate Certificate Authorities without changing their applications or sacrificing availability. I developed the backend service, lifecycle APIs, state model, and console workflows behind one stable issuance endpoint.

Zero-downtime rotationState-aware routingEnd-to-end validation

This portfolio describes the engineering problem at a high level and contains no internal code or proprietary architecture.

Production pool available
Single endpointroutes by member state
A
Authority 01Active · receiving traffic
B
Authority 02Draining · finishing requests
C
Authority 03Disabled · excluded
05
Omnidirectional Stereo CameraRaspberry Pi 5 · motors · encoders · 360-degree capture
06
Facial Expressions to EmojisOpenCV · four CNNs · 1,500+ images · 83% accuracy

02 / Experience

Shipping across cloud, autonomy, and AI.

01

Amazon Web Services

Software Developer Intern

May – August 2026Herndon, VA

Enterprise certificate customers needed one reliable interface for coordinating multiple stateful Certificate Authorities.

  • Developed a Java service and REST APIs that let customers manage Certificate Authority pools through one endpoint.
  • Defined active, disabled, and draining states. Added routing and failover for zero-downtime rotation.
  • Integrated AWS Fargate and console workflows, then verified recovery paths with RDE and end-to-end tests.
02

Amazon Web Services

Software Developer Intern

May – August 2025Herndon, VA

Certificate workflows needed stable references while the cloud resources behind them changed.

  • Implemented Java APIs and DynamoDB-backed alias resolution for persistent human-readable identifiers.
  • Integrated aliases with certificate issuance and revocation across new and existing APIs.
  • Wrote unit and integration tests for consistency, invalid references, error handling, and backward compatibility.
03

Forterra

Software Engineering Intern

May – August 2024Clarksburg, MD

An autonomous platform needed synchronized visual capture and interpretable sensor signals for downstream perception.

  • Coordinated eye-width cameras, motors, and encoders through Raspberry Pi 5 software for 360-degree stereoscopic capture.
  • Deployed the system on a General Dynamics robotic platform and validated repeatable capture behavior.
  • Created MATLAB signal-processing workflows that improved sensing accuracy by 35% and reduced noise by 20% against the raw-signal baseline.

03 / About

Product judgment backed by systems depth.

I enjoy engineering work where reliability is visible to the people using the product. That has taken me from cloud APIs and autonomous hardware to agent runtimes, retrieval systems, and GPU inference.

I’m pursuing a B.S. in Computer Science at Purdue University and expect to graduate in May 2027.

Languages

Python, TypeScript, JavaScript, Java, C++, C, CUDA, SQL, HTML/CSS

AI and systems

PyTorch, Hugging Face, vLLM, TensorRT-LLM, Triton, RAG, model routing, evaluation

Product and cloud

React, Node.js, GraphQL, REST APIs, AWS, DynamoDB, Fargate, Docker, Kubernetes