Skip to content
Portrait of Kevin Steverson

Kevin Steverson

Staff AI Engineer

I design and ship production AI systems: LLM platforms, real-time voice agents, and the evaluation and infrastructure that keep them reliable.

Atlanta, GA · Open to Staff & Principal roles

01

About

Eleven years turning machine learning into software people rely on.

I’m a Staff AI Engineer with eleven years of experience turning machine learning into software people rely on. Most recently at MILL5, I led the architecture of production AI platforms that combine large language models, classical ML, voice and autonomous agents — systems where latency, cost and correctness all matter at once.

I came to AI through backend engineering. I started on payment APIs at PayNearMe, moved into predictive analytics and NLP at Health Catalyst, and spent three years at WellSky building document intelligence and search for healthcare workflows. That path left me with a habit I still rely on: treat models like any other production dependency — measured, monitored and replaceable.

These days I spend most of my time on the less visible parts of AI engineering: evaluation, retrieval quality, model routing, and the standards that let several teams ship AI features without breaking each other’s work. I also enjoy mentoring engineers who are making the same move from software into machine learning.

Building production software, the last eight in ML and AI
11 yearsBuilding production software, the last eight in ML and AI
Lower inference cost from model routing and semantic caching
52%Lower inference cost from model routing and semantic caching
AI requests a day on platforms I designed
10k+AI requests a day on platforms I designed
Healthcare documents processed a year by my NLP pipelines
1.5M+Healthcare documents processed a year by my NLP pipelines

02

Experience

From payments APIs to Staff AI Engineer, across four companies.

  1. 2024 — 2026

    Staff AI Engineer · MILL5

    Led AI architecture across client platforms, from LLM orchestration and retrieval to real-time voice and multi-agent automation.

    • Architected production AI platforms combining LLMs, classical ML, voice AI and autonomous agents, reducing manual operational workload by 65%.
    • Built an evaluation framework on golden datasets and automated regression tests, cutting regression detection from 9 days to under 4 hours.
    • Introduced model routing and semantic caching that moved 70% of traffic to smaller models and lowered inference cost by 52%, with no measurable loss in quality.
    • Engineered real-time voice agents (ASR, LLM orchestration, TTS) with 45% lower average interaction latency.
    • Set AI engineering standards, architecture reviews and automated quality gates across 4 engineering workstreams.
    • LLMs
    • RAG
    • LangGraph
    • Voice AI
    • Python
    • FastAPI
    • Kubernetes
  2. 2021 — 2024

    Senior AI / Full-Stack Engineer · WellSky

    Built NLP, search and machine learning features inside healthcare software used for care coordination and operations.

    • Architected NLP and document intelligence pipelines processing 1.5M+ healthcare documents a year, reducing manual data extraction by 68%.
    • Built semantic search on embeddings and vector databases, improving retrieval relevance by 35%.
    • Shipped forecasting and classification models that were 24% more accurate than the existing baselines.
    • Redesigned asynchronous processing for 3.2× throughput and cut average API response times by 40%.
    • Python
    • PyTorch
    • Scikit-learn
    • NLP
    • Vector Search
    • Node.js
    • PostgreSQL
  3. 2018 — 2021

    Machine Learning / Software Engineer · Health Catalyst

    Moved from backend engineering into applied machine learning on a healthcare data and analytics platform.

    • Developed predictive and anomaly detection models that beat statistical baselines by 22% and improved detection precision by 27%.
    • Engineered data pipelines handling 5M+ records a month, reducing data preparation time by 55%.
    • Built NLP pipelines that turned unstructured documents into structured data, cutting manual processing by 48%.
    • Raised automated test coverage from 41% to 88%, contributing to a 61% year-over-year drop in production incidents.
    • Python
    • TensorFlow
    • Scikit-learn
    • Django
    • Flask
    • PostgreSQL
  4. 2015 — 2018

    Software Engineer · PayNearMe

    Backend and frontend engineering for a payments platform, where correctness and latency were not negotiable.

    • Built payment processing APIs and services handling 100,000+ transactions a month.
    • Reduced average transaction processing latency by 35% through service and query optimization.
    • Implemented fraud monitoring and risk checks that improved suspicious-activity detection by 26% and reduced processing errors by 32%.
    • Built reusable React components that cut feature development time by 30%.
    • APIs
    • Payments
    • Fraud Detection
    • React
    • Testing

Education

2010 — 2015

University of Arizona logo

B.S. in Computer Science

University of Arizona

View full resume

03

Selected work

Five systems I designed and shipped, written up as short case studies. Client names and confidential details are left out.

  1. An evaluation framework that catches LLM regressions in hoursGolden datasets, automated regression tests and quality metrics that turned “it feels worse” into a failing check.Regression detection from 9 days to under 4 hours · MILL5
  2. Model routing and semantic caching for LLM trafficSending each request to the smallest model that can handle it, and not paying twice for the same answer.52% lower inference cost, no measurable quality loss · MILL5
  3. Hybrid retrieval for more relevant RAG answersCombining keyword and vector search, reranking and contextual chunking so the model sees the right evidence.38% improvement in answer relevance · MILL5
  4. Real-time conversational voice agentsStreaming speech recognition, LLM orchestration and speech synthesis tuned for natural turn-taking.45% lower average voice interaction latency · MILL5
  5. Document intelligence for healthcare workflowsNLP pipelines that turn referrals, forms and clinical documents into structured, searchable data.1.5M+ documents a year, 68% less manual extraction · WellSky

04

Expertise

The tools and practices I use most, grouped by area.

Generative AI & Agents

  • LLMs — OpenAI, Claude, Gemini
  • RAG & hybrid search
  • LangChain, LangGraph, LlamaIndex
  • Pinecone, Weaviate, Chroma, pgvector
  • Fine-tuning & prompt engineering
  • Evaluation, guardrails & observability

Machine Learning

  • PyTorch, TensorFlow
  • Scikit-learn, XGBoost
  • Predictive modeling & forecasting
  • Recommendation systems
  • NLP & computer vision
  • Feature engineering & model evaluation

Voice & Speech

  • ASR, speech-to-text & text-to-speech
  • Real-time conversational agents
  • Voice activity detection
  • Streaming audio & WebRTC
  • Whisper, Deepgram, ElevenLabs
  • Twilio Voice & telephony

Backend & Full-Stack

  • Python — FastAPI, Django, Flask
  • Node.js, TypeScript
  • React, Next.js, Vue.js
  • REST, GraphQL, WebSockets
  • PostgreSQL, MongoDB, Redis
  • Event-driven microservices

Cloud & MLOps

  • AWS, Google Cloud, Azure
  • Docker, Kubernetes, Terraform
  • MLflow, Weights & Biases
  • CI/CD with GitHub Actions
  • Langfuse, Sentry, OpenTelemetry
  • Data pipelines & orchestration

Leadership

  • AI architecture & system design
  • Architecture reviews
  • Engineering standards & quality gates
  • Mentoring engineers into ML
  • Concept-to-production delivery
  • Performance engineering

05

How I work

Five principles I bring to every AI project.

  1. Measure before you optimize.

    Every AI feature I ship starts with an evaluation set. If we can’t say whether a change made things better, we aren’t ready to make it.

  2. Use the smallest model that does the job.

    Large models are a good place to start and an expensive place to stay. Routing and caching are often the cheapest performance work available.

  3. Latency is part of the product.

    In voice especially, a correct answer that arrives late is still a bad answer. I budget latency per stage the same way I budget cost.

  4. Treat models as dependencies.

    Versioned, monitored, observable and replaceable. Prompts and providers change; the system around them shouldn’t have to.

  5. Leave teams faster than you found them.

    Standards, reviews and quality gates are only worth it when they make it easier for other people to ship.

06

Contact

Reach me by email or phone, or download the full resume.

I’m focused on agentic systems and evaluation — making multi-step AI workflows observable and predictable enough to trust with real business processes. I’m open to conversations about Staff and Principal AI engineering roles, and to advising teams taking an AI product from prototype to production.

Email is the best way to reach me. If you’re hiring, a short note about the team and the problem you’re working on is plenty to start with.

Location
Atlanta, GA
Email me