Dylan
Hampton

I build LLM platforms inside regulated cloud environments, and AI workflows that business teams use on their real work. When a result misses, I trace why and say so.

RoleSoftware engineer, applied AI
Based inSt. Peters, Missouri
Open toRemote or relocation
CitizenshipU.S.
EducationB.S. Computer Science
Page revisedNovember 16, 2026

Selected work

Employer and client code is private, so this page describes the work and draws how it fits together. Red notes mark the limits.

Stauder Technologies, 2026

LLM gateway with per-user attribution

Problem
About 30 engineers share Amazon Bedrock in an ITAR-controlled AWS GovCloud environment, and usage has to be traceable to a person and a project.
What I built
A LiteLLM gateway in front of Bedrock, plus a companion service that stamps each request with user and project, so usage shows up in AWS's own logs. Identity comes from the authenticated key, not headers a client could fake. Only metadata is attached, never prompt content.
What I found
The obvious approach failed because request-body metadata is stripped on the /v1/messages route. I traced that and fixed it with a pre-call hook. I also designed and directed a FastAPI bridge (AI-assisted implementation) so Codex CLI traffic goes through the same gateway.

LiteLLM, Amazon Bedrock, AWS GovCloud, FastAPI, CloudWatch, Athena

Engineers' AI toolssigned in with a per-user key
Codex CLIthrough a FastAPI bridge
LiteLLM gatewaypre-call hook reads identity from the keyuser + project attached
Amazon Bedrock, GovCloudrequest metadata travels with the call
AWS's own logsinvocation logs, CloudWatch, Athena
Request path. Only identity metadata is attached. Prompt content never goes into the logs.

Stauder Technologies, 2026

AI rollout for an engineering team

Problem
Engineers working with controlled unclassified information needed AI tools they could use without breaking data-handling rules.
What I built
Bedrock Guardrails and usage policies supporting FedRAMP High. An onboarding path with bootstrap scripts, data-handling rules and a shared prompt library, then hands-on workshops. I rolled out Codex CLI on Windows after finding a PowerShell 5.1 incompatibility, and took part in model selection and migration planning for a multi-provider gateway.
Result
The whole team of about 30 now uses it every week.

There are no formal productivity metrics yet.

Bedrock Guardrails, FedRAMP High, Codex CLI, PowerShell, workshops

MarkSet (AI workflow consultancy), October 2026

Claude workshop and skills for a mechanical contractor

Problem
A contractor's project managers and estimators wanted to use Claude on their own jobs, which live in SharePoint and Outlook.
What I built
A paid 90-minute workshop on their live jobs. I cut it from a 2-hour lecture into a hands-on format and ran a dry run with a real PM first. Then three Claude skills in their Enterprise org: a job briefing that pulls each job's files and mail into one page, a morning report across active jobs and open bids, and a writing-voice profile.
What I found
The dry run caught the briefing stating things that weren't true, so every line now carries a tag naming its source. The SharePoint connector silently stops at about row 1,950, so the report reads a local copy with a row checkpoint. After the workshop, the owner asked for a company-wide knowledge system.

The knowledge system is designed, not built.

Claude Enterprise, Claude skills, SharePoint, Outlook

Job briefing

Duct submittal came back with comments.Email
Crew of six on site Tuesday.Log
Owner wants the ceiling closed before the holiday.PM
The roof-unit schedule looks tight.My read
Evidence tags. Made-up lines, not client data. Each tag says whether a line came from the daily log, an email, the PM, or the model's own judgment, so a reader can check it.

MarkSet, 2026

Estimating agent tested against completed bids

Problem
A construction estimate from an AI model can look right and still be far off. I wanted to know how far, using jobs where the award is already known.
What I built
An estimating agent, scored against completed bids.
What I found
Within 0.3% of the award on a healthcare build. A 36.6% miss on a lab renovation, which I traced to missing calibration for that job type. In a later test, a plain model with no estimating guidance mispriced the same lab job by about 3x.

Two jobs is a small sample, and I treat it as one.

Healthcare build, agent0.3% off
Lab renovation, agent36.6% off
Lab renovation, plain modelabout 3x off
0%20%40%
Distance from the awarded price. The third bar runs past the scale.

Personal project, October 2026 to now

Prediction-market agent

Problem
An agent that can place trades needs its limits enforced in code, not in its prompt.
What I built
An agent that scans every open Kalshi market (5,000+) on a schedule. Spending limits live in code. Small actions run on their own, and judgment calls go to a human approval inbox that records the reason for each decision. 160+ tests run against a simulated exchange.

It is running a paper month before any real money, so there are no returns to report.

Python, scheduled runs, simulated exchange

Scan every open market5,000+, on a schedule
Spending limits, checked in codea model can't change them
Small actionruns on its own (paper mode for now)
Judgment callwaits in the approval inbox, reason recorded
Decision path. Limits sit in code, ahead of any model output.

Personal project, June 2026 to now

LearnFrame

Purpose
A meta-system that sets up purpose-built AI systems and keeps a shared library of patterns for them.
What I built
A repo-level agent setup I direct daily: context files, 12 shared skills, hooks, and a nightly self-improvement job that has run since June 2026. It has set up and drives ten separate projects. Built with AI coding agents under my direction (430+ commits).
The rule
A pattern enters the shared library only if it beats a plain-model baseline in a scored head-to-head test.

Claude Code, Codex, hooks, scheduled jobs

Candidate patterna skill or lesson from one project
Scored head-to-headsame task, with the pattern and with a plain model
Beats the baselineenters the library
Doesn'tstays out
Entry rule. The library grows only through this test.

Projects and research

Smaller builds and measurements. I design, direct and test these, and AI coding agents write most of the code.

Research, 2026

Where an agent's context goes

I measured how much of Claude Code's context its built-in tool schemas use before any work starts: about 34,800 tokens, or 17.4% of a 200,000-token window.

Context Headroom is a method I designed and published as a LinkedIn post: compact early, at around 40% of the window, into a verified ledger, instead of waiting until the window is nearly full.

I haven't collected usage data on Context Headroom yet.

17.4%
Compact here, ~40% 200k
One 200,000-token window. The hatched part is used by tool definitions before the first message.

Personal project, July 2026

On-device phone agent

~16 tokens/sec, measured on the phone

llama.cpp serving Gemma 4 E2B on a Galaxy S25 Ultra through Termux, with phone tools through Termux:API. The shell is allowlisted, and any action with side effects waits for me to type "yes". I ran and debugged it on the device; Claude wrote the script.

Personal project

ICM agent package

A standard-library Python agent with grammar-constrained decoding and a git-backed knowledge tree that records where each fact came from. 16 of 16 tests pass.

So far the tests run only against a stand-in model, not a real one.

Personal project, November 2026 to now

Personal news poster

My daily market and tech brief. It collects from a public list of sources every four hours and posts a visual page each weekday morning, plus a Saturday wrap. It informs only and never makes trade calls. To earn its keep, it has to beat a plain prompt in a blind week-long head-to-head, on at least 4 of 5 days.

It went live November 14, 2026. The head-to-head week hasn't run yet.

Personal project

Parametric 3D-printed parts

Functional parts I model as code, print and measure. Fit clearances come from a printed test coupon rather than guesses, and a part counts as done only when it has been measured against what it had to do.

On the side, I'm designing a mobile factory-automation puzzle game: one fair daily puzzle for everyone, plus a base that grows between dailies. Coding agents build it from my design.

Experience

Software Engineer, Stauder Technologies

St. Peters, MO. May 2024 to present.

  • The AI platform work above.
  • Message encode/decode for VMF (MIL-STD-6017), Link 16 (MIL-STD-6016) and JREAP-C (MIL-STD-3011) across two client applications, with gateway logic translating between standards and revisions, including Cursor on Target. Validated conformance against each standard.

Founder, MarkSet (AI workflow consultancy)

St. Louis, MO. April 2026 to present.

  • The workshop, skills and estimating test above.
  • A skill that turns a prospect's company URL into a fit analysis, workflow maps and a working demo dashboard.
  • A client data guide for customers that restrict AI use: where data goes, what the model provider keeps, and a per-job AI rule.

B.S., Computer Science

Missouri University of Science and Technology, May 2024.

Skills

AI engineering
LLM APIs and tool use, Claude skills and hooks, agents, LiteLLM, Amazon Bedrock and Guardrails, evals against real completed work, directing Claude Code and Codex, local inference with llama.cpp
Safety and security
ITAR and CUI environments, FedRAMP High, DoD IL4/IL5, identity from authentication, spending limits in code, human approval before side effects, audit logging
Integration
Standards-based message encode/decode and translation, conformance validation, REST APIs (FastAPI, Java HttpServer), SharePoint and Outlook connectors
Languages and tools
Python, TypeScript/JavaScript, Java, C++, C#, Bash, PowerShell, AWS (GovCloud, Bedrock, IAM, CloudWatch, Athena), Docker, Git, GitLab, Linux, SQLite, MongoDB