Qosmit LLC · Independent consultancy

Modernize the infrastructure.Then make it think.

I modernize systems that already carry real work — legacy servers, hand-managed deploys, brittle pipelines — without treating production like a clean-room rewrite. Then I build AI into the places where it can be measured, reviewed, and worth operating.

Start a conversation See the work

Years in site reliability and production engineering

11+

Tenant environments migrated to containers and running in production

Engineer on your project — you talk to the person building it

Where I've done the work

MeijerCorewell HealthCompose LLCCrystal Clean Auto Detailing

From enterprise platforms to independently operated client systems.

Services

Two things, done properly.

I don't offer a menu of everything. I do infrastructure work and I do AI work, and the reason they sit together is that the second one only pays off once the first is solid.

Cloud & DevOps modernization

Legacy infrastructure turned into something a small team can actually operate — versioned, repeatable, and cheaper to run than what it replaced.

Legacy-to-container migration

Windows/IIS and VM-per-customer estates moved onto ECS Fargate or equivalent. One image, many tenants, differences expressed as configuration instead of snowflake servers.

Infrastructure as code

Terraform and OpenTofu with a clear ownership line between what the pipeline controls and what the state file controls, so a plan never surprises you.

CI/CD you can trust

GitHub Actions pipelines with real promotion gates — auto-deploy to test, human-gated releases, staged rollout waves, and a rollback path that has been rehearsed.

Cost and reliability review

Right-sizing, Spot for non-production, and finding the orphaned storage and stopped-but-still-billing resources that quietly accumulate for years.

Observability

OpenTelemetry, Prometheus, Loki, Tempo, and Grafana wired into one operating view, with authenticated ingestion and enough context for an alert to start with a hypothesis.

Secrets and least privilege

Credentials in a parameter store rather than a config file, deploy roles scoped to what they actually need, and permission gaps caught before the apply rather than halfway through it.

AI & agentic systems

LLM work grounded in evaluation rather than demos. Correctness, cost, and escalation paths are part of the system — not details added after the prompt works.

Agent pipelines that do real work

Builder, reviewer, and escalation loops applied to concrete tasks, with isolated context, bounded retries, and independent verification against the actual work product.

Evaluation harnesses

Benchmarking infrastructure that separates convergence from correctness and measures token use, cache behavior, elapsed time, cost, and human-review requirements.

AI-assisted operations

Alerts routed through an analysis step that triages the incident and opens a draft pull request with a proposed fix, so on-call starts from a hypothesis.

LLM features in production apps

Claude and OpenAI integrations built into real products — streaming, tool use, context management, and cost controls that survive contact with users.

Practical AI strategy

An honest read on what to automate, what needs a human decision gate, and how to route routine work to economical models without hiding high-risk judgment inside a fallback.

The data plumbing underneath

Most failed AI projects are retrieval projects in disguise. Getting the right context to the model — indexed, permissioned, and fresh — is usually the actual work.

Selected work

Systems that were already running.

Most of my work is on platforms that customers were using the whole time. The interesting constraint is never the architecture — it's doing it without a blackout.

Crystal Clean Auto Detailing

Contract · Cloud modernization

A per-customer Windows fleet became one container image

The situation

Every dealership ran its own Windows EC2 instance with IIS, deployed by hand over WinRM from a bastion host. Each new customer meant another server to build, patch, and remember.

What I did

I rebuilt the platform on ECS Fargate: a single .NET image deployed to one isolated container per dealership, each with its own database, differentiated only by environment variables injected from Terraform and SSM. Subdomain routing through an ALB, secrets never in plain text, non-production on Spot capacity.

Result

Onboarding a new dealership went from building a server to a five-step runbook plus a Terraform apply.
A release pipeline with an automatic test environment, a human-gated client sign-off environment, and staged production waves that refuse to run out of order.
Found and retired a fleet of stopped legacy instances still billing for attached storage every month.

AWS ECS Fargate
Terraform
GitHub Actions
.NET / EF Core
RDS MySQL
ALB
SSM

Compose LLC

Co-founder · Senior DevOps Engineer

Scaling the platform while rebuilding its delivery foundation

The situation

Compose was operating a platform serving more than 50,000 users while carrying a legacy .NET 4.6.1 backend and a delivery process that needed to support multiple environments safely.

What I did

As co-founder and Senior DevOps Engineer, I launched, monitored, and maintained the platform, built AWS CI/CD pipelines with production approval controls, and helped rebuild the backend on Node.js and PostgreSQL.

Result

Supported a production platform serving 50,000+ users.
Reduced backend call response times by 80% through the .NET-to-Node.js and PostgreSQL rebuild.
Established multi-environment deployments with an explicit production approval gate.

AWS
CI/CD
Node.js
PostgreSQL
.NET 4.6.1
Observability

Corewell Health

Senior Site Reliability Engineer · FinOps

Automating AWS invoice reconciliation into seven accountable cost centers

The situation

AWS invoice data did not map cleanly to business ownership. Recurring reconciliation work was difficult to scale, leadership lacked a simple top-down view of cloud costs, and significant portions of spend could not be readily explained.

What I did

I designed and automated a reconciliation process that maps AWS invoice units into seven cost centers, creating a repeatable ownership model for organization-wide cloud spend. I then used that clearer cost view to investigate and resolve significant previously unknown AWS charges.

Result

Consolidated cloud cost ownership into seven accountable cost centers.
Replaced recurring manual invoice reconciliation with a repeatable automated workflow.
Gave leadership a top-down view of AWS costs and resolved significant previously unexplained spend.

AWS billing
FinOps automation
Cost allocation
AWS Organizations
Cloud cost governance

Meijer

Site Reliability Engineer · Store Customer Experience

Reliability engineering for a new distributed retail system

The situation

A new distributed customer-experience system needed measurable reliability, useful operational visibility, and a support model that did not depend on repetitive manual intervention.

What I did

I architected and led the SRE monitoring and alerting work, established a 99.9% SLO through root-cause analysis and cross-functional collaboration, automated Level 1 processes, and built most of the Dynatrace dashboards used by stakeholders and IT.

Result

Reduced manual intervention and cut $70,000 in annual operating costs.
Delivered 80% of the Dynatrace dashboards used to understand the distributed system.
Gave development, architecture, and operations teams a shared reliability target and operating view.

Dynatrace
SLOs
Distributed systems
PowerShell
Selenium
Groovy

The lab

The questions I turn into working systems

Consulting work is often private. These projects are the clearest sample of how I approach operations, authorization, agent evaluation, and evidence-driven product development.

ObservaStack

A self-hosted observability platform that brings metrics, logs, traces, alerting, agent registration, and authenticated telemetry ingestion together. Its AI-assisted triage path can turn a firing alert into a concrete hypothesis and draft fix.

OpenTelemetry
Grafana
AI-assisted operations

planpermit

An AWS authorization analyzer designed to explain effective access in plain language: what an identity can do, which policy layer blocks it, which roles it can reach, and where redundant or privilege-escalating paths exist.

AWS IAM
Identity Center
Access graphs

Aloop

A benchmarking harness for multi-agent coding pipelines. It compares local and cloud builders on ground-truth tasks while reconciling per-stage telemetry — because a loop that reaches PASS is not necessarily a loop that got the answer right.

Agent evaluation
Telemetry
Local models

Stock Edge

A paper-trading research platform built to test whether public insider activity contains a durable signal. The project emphasizes reproducible datasets, point-in-time correctness, simple baselines, realistic execution, and explicit human review before a backtest becomes a claim.

Research systems
Backtesting
Evidence gates

Approach

How I work

One engineer, direct line, no account manager in between.

Read the system first

Before proposing anything I want to see the infrastructure, the deploy path, and the parts everyone is afraid to touch. Most engagements start with a week of reading and one honest document about what's actually wrong.

Migrate without a blackout

Legacy and new run side by side until the new path has earned the traffic. Cutovers happen per tenant, per environment, in waves — never as a single weekend with no way back.

Leave it documented

Runbooks, architecture decisions, and the reasoning behind the non-obvious choices. The goal is that your team can operate it after I'm gone, not that you need me on retainer.

Test the claim, not just the code

Passing tests can prove that software behaves as specified; it cannot prove that a model, strategy, or research conclusion is sound. I define baselines, ground truth, and human decision gates wherever judgment still matters.

About

One engineer. No layers in between.

Qosmit LLC is the consulting practice of Hunter Dacey, a Senior Site Reliability Engineer based in Grand Rapids, Michigan. Since starting in site reliability in 2018, I've built and operated systems inside a national retailer, a large regulated health system, and a software company I helped found — and I now take that experience to teams that need senior engineering without a full-time hire.

The through-line in my work is unglamorous: making systems legible. Most of what slows a company down isn't a missing feature, it's infrastructure nobody fully understands anymore. I like being the person who maps it, moves it somewhere sane, and writes down how it works.

That same instinct shapes my AI work. I care less about a convincing demo than whether the system exposes its assumptions, measures the right outcomes, and knows when to hand a decision back to a person.

I work with a small number of clients at a time, directly, on engagements where the outcome is measurable.

AWS
AWS IAM
AWS Organizations
Terraform
OpenTofu
ECS Fargate
Docker
GitHub Actions
Next.js
TypeScript
React
.NET
C#
Python
Node.js
MySQL
PostgreSQL
Prisma
Entity Framework
OpenTelemetry
Prometheus
Grafana
Loki
Tempo
Claude API
OpenAI API
Local models
Agent orchestration
Evaluation harnesses

EntityQosmit LLC · Michigan

Founded2025

EngagementsProject-based and fractional

AvailabilityTaking new work

Start a conversation

Let's talk about what you're running

Tell me what the system is, what it's costing you, and what you wish it did instead. If I'm not the right fit I'll say so quickly.