Case studies

Selected work.

Current and recent engagements, described the way we run them. Client names are withheld and details generalized; the numbers are real and shared with permission. References available in scoping conversations.

Modernization & AI-assisted migration Education-services company

AI-accelerated SQL Server → MySQL reporting modernization

Situation

A reporting estate built over many years on SQL Server — including stored procedures, SSRS reports, subscriptions, schedules, SSIS packages, and SQL Agent jobs — needed to move to MySQL 8 without changing the business meaning or output of hundreds of reports. The scale and dependency complexity made a manual rewrite approach too slow, inconsistent, and difficult to validate safely.

Approach

Built an AI-assisted migration factory that inventories stored procedures, functions, views, reports, packages, schedules, and downstream dependencies across the reporting estate.
Used specialized AI agents to analyze source logic, extract business intent, map dependencies, generate MySQL-compatible rewrites, identify unsupported SQL Server patterns, and surface high-risk conversion decisions for human review.
Established a governed lifecycle for every object: discovery, dependency resolution, intent extraction, rewrite, validation, exception review, approval, and release.
Created a dedicated QA harness that generates representative test cases and compares SQL Server and MySQL results row by row, preventing syntactically correct but functionally wrong rewrites from being approved.
Applied human approval gates to parity exceptions, architectural changes, and production-impacting decisions rather than allowing autonomous promotion.
Extended the migration beyond stored procedures to include SSRS delivery, subscriptions, SSIS dependencies, SQL Agent workflows, environment synchronization, and cutover planning.

Where it stands

AI transformed the effort from a sequence of manual rewrites into a controlled modernization program. A prioritized queue of more than 300 reports and procedures now moves through a repeatable factory cadence, with business intent preserved, dependencies documented, and output parity proven before approval. The result is faster delivery without sacrificing engineering control: AI performs the high-volume analysis and conversion work, while experienced database oversight governs architecture, risk, validation, and final acceptance.

Verification
Object counts come from the migration factory's tracking inventory; parity is proven per object by a harness comparing source and target result sets row by row, with human sign-off gating approval.

Limitation
The migration is in progress — figures describe the working queue and validated objects, not a completed cutover.

Database infrastructure National logistics provider

End-to-end SQL infrastructure management

Situation

A growing SQL Server estate spanning production, staging, QA, and development required consistent engineering standards across infrastructure design, buildout, operations, security, performance, and lifecycle management.

Approach

Own the SQL infrastructure lifecycle from capacity planning and technical specifications through server buildout, Availability Group design, database deployment, monitoring, backup, recovery, and ongoing administration.
Apply repeatable operating standards for staging refreshes, SQL Agent job health, data sanitization, access control, orphaned-user remediation, validation, documentation, and rollback.
Continuously tune queries, indexes, storage layouts, maintenance routines, and high-churn tables to improve performance and reduce operational overhead.
Adopt and enforce security and management best practices across permissions, service accounts, production access, environment separation, change control, and recovery readiness.

Where it stands

The SQL estate is managed as a controlled platform rather than a collection of individual servers: infrastructure is built to documented standards, recurring operations are automated and review-gated, failures are surfaced with clear ownership, and performance, security, and recovery practices are applied consistently across environments.

Verification
Refresh runs are driven by control tables and end with a validation phase and a written report; job health is tracked in dated remediation reports; and the partitioning work's Query Store-based impact report measured the nightly maintenance window falling from roughly 42 minutes to 12.

Limitation
Figures describe the estate as currently operated — the maintenance-window measurement is from its first full cycle, and job-failure counts vary week to week.

Adjacent engineering

The same discipline — evidence, gates, human review — applied beyond the database estate.

Reporting & data automation Real-estate financing advisory

Deal-sourcing signal engine — validated before it was built

Situation

An advisory working in federal tax-credit financing (NMTC) asked whether an automated "signal engine" could surface manufacturing expansion projects earlier than the usual intermediaries — and whether building one was within reach.

Approach

Instead of building first, we ran the idea through a blind, reject-by-default adversarial panel — three independent refuters, each verifying the idea's load-bearing claims against live public sources: real county board agendas, permit-notice feeds, a live geocoder-to-census-tract call, and commercial vendor pricing.
Round one rejected the premise 3-for-3: the early public signals are systematically anonymized, the complete ones publish only after deals are secured, and the detection layer turned out to be a purchasable commodity feed.
We reframed the commercially real window — post-announcement, pre-financial-close, with a likely tax-credit-fillable gap — and rewrote the plan as a gated sequence: retrospective backtest, manual conversion pilot on purchased feeds, and a build limited to the differentiated screening overlay no vendor sells.
The retrospective backtest is complete, and the eventual build is specified with its limits declared: evidence-preserving LLM extraction with human QA, indeterminate geography as a first-class state, human-confirmed event matching, and a review queue — no silent automation.

Where it stands

The client avoided building a product whose original premise was demonstrably false. The adopted plan prices every remaining unknown at a gate that can kill the project for under $1,000 — and if the gates pass, the build covers only what commodity vendors don't sell.

Verification
Every refuter verdict is documented with per-claim evidence — live checks of the public sources, a live geocoder-to-tract call, and verified vendor pricing — and the adopted plan declares its gates in writing.

Limitation
Validation and the retrospective backtest are complete; the conversion pilot — the gate that decides whether anything gets built — is still ahead.

Reporting & data automation Real-estate operator

Automated content and distribution pipeline

Situation

A solo operator needed a steady weekly content presence across social networks without hiring a team or spending hours a day posting manually.

Approach

Built a scheduled pipeline on AWS: AI generation produces scripts, captions, and rendered vertical video on a weekly cadence, and a daily scheduler distributes finished posts across networks through a scheduling API.
Everything is infrastructure-as-code with CI checks, so the system is reproducible and changes are reviewed before deploy.
Per-market configuration lives in data, not code — expanding to a new market is a config row, not a rewrite.

Where it stands

A hands-off weekly production and daily distribution cadence running on tens of dollars a month of cloud infrastructure, operated by one person.

Verification
Infrastructure cost is read directly from the cloud bill; generation and posting runs are scheduled, logged, and reviewed through the project's CI pipeline.

Limitation
The system measures output cadence and cost, not audience outcomes — distribution results accrue on the client's channels, not in this pipeline.

When clients engage us

A business-critical database environment with recurring incidents, performance problems, or operational risk
A SQL Server, MySQL, PostgreSQL, reporting, or cloud migration that requires dependency discovery, validation, and controlled cutover
Reporting, ETL, SQL Agent, backup, recovery, or environment-refresh processes that are fragile, manual, or poorly owned
A need to design, build, standardize, or manage SQL infrastructure across production and lower environments
An AI, data-product, or automation idea that should be validated before significant development spending
A repeatable content or operational workflow that can be governed, automated, monitored, and scaled
Limited internal senior database capacity for a high-risk initiative
A need for an independent technical assessment before a major architecture, migration, or investment decision

When another provider is a better fit

General IT help desk or desktop support
Commodity website or mobile-app development
Generic chatbot projects without a defined data, workflow, or business problem
Standalone social-media management or creative marketing services
Experimental AI projects without measurable outcomes, usable data, or human review controls
Low-cost staff augmentation without clear ownership, scope, and acceptance criteria
Projects that require unsupported claims, fabricated metrics, or uncontrolled automation
Routine implementation work where the solution is already known and no senior engineering judgment is required

Not sure whether the engagement fits? Use the initial consultation to define the problem, determine whether specialist involvement is warranted, and identify the smallest responsible next step.

Discuss a database problem See how engagements are scoped