CTO & Head of Applied AI · Agentic Systems · Predictive Machine Learning · LLM Engineering

I architect and build applied AI systems: conversational agents, autonomous engineering workflows, predictive models, and the software that makes probabilistic behavior reliable in production.

I'm CTO at Snap eHealth, where I lead AI engineering across healthcare applications and stay hands-on with system architecture, implementation, evaluation, debugging, and delivery. My focus is not simply adding an LLM to a product. It's designing the complete system around the model: tools, data, state, verification, security, human oversight, and measurable outcomes.

My work spans Python, scikit-learn, TypeScript, Go, APIs, distributed systems, AWS, and Kubernetes. I'm especially interested in combining traditional predictive ML with LLM-powered agents so each component does the work it is best suited to do.

At Snap, I lead engineering for Sally, an LLM-powered healthcare voice agent with real-time conversation, tool integration, background analysis, and structured visit-brief extraction. My work includes the architecture and production hardening required for patient-facing AI: consent and authorization boundaries, deterministic application controls, structured outputs, scenario-based evaluations, and failure handling.

I also architect autonomous software engineering systems using Codex, Symphony, and specialized agents for implementation, testing, review, debugging, and remediation. The engineering challenge is not generating code; it's controlling what agents can do and establishing evidence that their work is safe to accept.

Key design concerns include:

• Orchestration and tool use: agent handoffs, scoped permissions, isolated execution, and controlled access to external systems.
• State and reliability: durable execution, cancellation, recovery paths, and bounded autonomous behavior.
• Evaluation: structured evidence, scenario tests, deterministic verification gates, and traceable results.
• Safety and governance: cost limits, auditability, human approval, and explicit authority over code merges and production changes.

I'm developing a Physician Referral Success Predictor to estimate whether a referral will result in a completed appointment within 30 days. It uses Python and scikit-learn and is being built with synthetic healthcare data for reproducible, privacy-conscious experimentation.

The project is focused on the full supervised-learning workflow rather than a model demo alone:

• Defining the prediction target and separating information available at referral time from future outcomes.
• Engineering features from referral, provider, and scheduling signals while checking for target leakage.
• Establishing an interpretable logistic-regression baseline and comparing it with more flexible classifiers as development progresses.
• Evaluating precision, recall, discrimination, calibration, and error patterns rather than reporting accuracy alone.
• Designing an inference API and examining how model probabilities can support — but not replace — deterministic eligibility and workflow rules.

Status: active development; this is not represented as a production-deployed or clinically validated model.

I'm also interested in how predictive models become tools within agentic workflows: a trained model computes a probability, an agent retrieves context and coordinates tools, and deterministic policies govern what can happen next.

I have led AI-assisted engineering and reliability work including Amazon Bedrock LLM evaluations, goldset reference datasets, guardrails, and an AI-driven test-intelligence system that turns product stories and defects into test requirements, API tests, and targeted retests.

I've worked on specialized QA and SRE agents, structured findings, deterministic triage, verification gates, and human escalation. These systems reflect an engineering principle I apply across AI products:

Models can propose and reason. Reliable software must still verify, constrain, observe, and account for what happens.

Much of my healthcare and autonomous-engineering work is proprietary to Snap; these descriptions are architectural overviews, not links to company source code. I use independent projects to demonstrate methods and implementation decisions publicly.

My public engineering work is increasingly centered on two complementary areas:

1. Predictive ML with defensible evaluation: train/test design, interpretable baselines, model comparison, probability calibration, and trustworthy inference.
2. Stateful agentic applications: graph-based workflows, tool calling, memory and context, evaluation harnesses, recovery, and human-in-the-loop controls. I'm exploring LangChain and LangGraph for this work, without treating framework adoption as a substitute for sound system design.

The longer-term goal is to connect them: predictive models provide evidence; agents coordinate actions; deterministic systems enforce constraints.

I have more than a decade of engineering leadership experience and remain a hands-on builder. At Snap, I've led delivery across enterprise healthcare products and AI applications. Earlier, I built a large-scale automated test platform with 2,300+ tests, compressing a roughly three-week regression process into approximately 30 minutes.

My engineering background includes distributed systems, APIs, cloud infrastructure, and production reliability. I apply that foundation to building, evaluating, and operating AI systems—with a particular focus on agentic applications and predictive machine learning.

I write about the systems I'm building and the engineering decisions behind them: LLM architectures, agent orchestration, AI evaluation, predictive modeling, failure modes, production safeguards, and technical leadership.

Website: thomashoward.dev

LinkedIn · GitHub · Website