How We Work
AI-accelerated. Engineer-led.
Our engineers use AI to move faster and stay accountable for everything that ships.
Whether we are modernizing defense training software or building an AI system, we work the same way: small loops, automated checks, and a person reviewing the result.
Our Method
Plan, Act, Observe, Evaluate
One method on every program, whether the work is done by our engineers, an AI agent, or both. Agentic AI is only as good as the loop around it, so we design that loop: what gets done, how it is checked, and where a person decides.
-
Plan
Scope the work and write the plan. Nothing proceeds until a person approves it.
-
Act
Build in an isolated workspace, with scoped permissions and allow-listed commands for any AI agent.
-
Observe
Run the tests, read the logs and check the results, inside your boundary.
-
Evaluate
Score the result against tests and evaluation harnesses, then a person reviews the diff.
Around every loop: guardrails, budgets, audit trails and human-in-the-loop review checkpoints.
Six Principles Behind Every Loop
Senior Engineers, Embedded
Experienced senior engineers join your program’s teams and cadence, work in your environments, and own outcomes rather than tickets.
Every Change Is Reviewable
Work arrives as a diff or pull request that a person reads and approves, whether a person or an AI agent wrote it. Plans stay read-only until approved, and nothing merges on its own.
Tests Before Trust
A change is not done until a test proves it, and the test fails without the fix. Pipelines run the tests and security scans on every change.
Guardrails Set Up Front
Tool permissions, allow-listed commands and spending limits are set before an AI agent runs, with an audit trail of what it did.
Your Boundary, Your Data
Models run in environments you control, so prompts, code and outputs stay inside your boundary. AI answers cite their sources, so people can check them.
Modernize in Place, Hand Off Cleanly
We refactor step by step, with infrastructure as code and observability from day one, and leave documentation and runbooks so the system outlasts the engagement.
We Use What We Ship
Our Delivery Is Powered by Topsoil
Our engineers use Topsoil every day. Backlog items are worked in an isolated git worktree and come back as a pull request for a person to review; Topsoil never merges on its own. Monthly account and user budgets are checked before every model call.
“For the past year, we’ve been building Topsoil and using it every day across Thazen’s own engineering teams. It changed how we deliver.”
— Zenon Rawley, CTO, Partner & Co-Founder, Thazen
Explore Topsoil (opens in a new tab)The Topsoil Story
2024
Bedrock Prototype
Our NAWCTSD Prize Challenge prototype uses Amazon Bedrock, RAG and AWS GovCloud (US) to help evaluators review proposals, with citations back to the source. It reaches the semi-finalist round.2025
AI in Our Own Delivery
We start building an AI coding agent for our own engineers, with the same method we apply to customer work: read-only plans, diffs, tests that must fail first, and a person on every merge.Oct 2026
Topsoil Ships
After a year of daily use across Thazen’s own engineering teams, we release Topsoil for customers and partners in the FedGov and DoW space.
Ways to Work With Us
Two Ways to Start
Embedded Program Engineering
Senior AI, cloud, DevSecOps and Palantir engineers working inside your program team as a subcontractor, on your backlog and in your environments, under your program’s security and delivery rules.
Focused AI Pilot
A scoped AI build against success criteria you set, in an environment you control.
Start With One Loop
Pick a backlog, a legacy component or an AI idea. We will scope a first loop with you.