AI and workflow systems I've actually built.
Practical buyer tools, automation assistants, and human-in-the-loop workflows, built around real business friction.
I build small, useful systems around the friction businesses live with: buyer questions that never stop, quote-prep bottlenecks, scattered operating data, repetitive review work, and go-to-market workflows that need structure before they need scale. Some use LLMs. Some use rules, data, and plain automation. The common thread is simple: the tool gives a person a clearer answer, a cleaner handoff, or fewer hours of repetitive work.
Built by Trey Sparks
What this work proves
I find the real workflow problem.
The useful automation is rarely the first thing someone asks for. I look for the repeated question, the bottleneck, the handoff, or the decision that is quietly costing time.
I turn messy judgment into structured logic.
Inputs, assumptions, edge cases, confidence levels, summaries, next steps. The parts a usable tool needs before anyone worries about the interface.
I build things nontechnical teams can use.
A tool should make the work clearer, not create a second job called operating the tool.
I keep humans in the loop.
The goal is not to replace judgment. It is to clear the repetitive grind so judgment gets spent where it matters.
I ship small before promising to transform anything.
No eighteen-month roadmap required. Start with the tool that answers one real question.
A product site, built, shipped, and instrumented end to end.
The problem. A manufacturer's products were buried on a slow, low-authority corporate site. Buyers could not scope a product, compare options, or start an inquiry without picking up the phone, and sales kept fielding the same questions.
What I built. A 31-page product-authority website, shipped to production and wired to its own analytics and lead capture.
Full product, industry, and how-it-works page sets, plus resource and tool libraries. Two working interactive decision tools shipped in production: a product-selection advisor and a configurator that turn buyer inputs into a quote-ready summary. An automated indexing pipeline: every code push deploys the site and submits new or changed pages to search engines automatically, so nothing waits on a slow crawl. Connected analytics (GA4, Google Search Console, Bing Webmaster Tools) and form-based lead capture. A scheduled weekly scan that checks indexing, search traction, and page issues, and emails a digest for review.
End-to-end GTM engineering: design, build, deploy, instrument, monitor. Interactive buyer-decision tools running in production, not mockups. Marketing and SEO automation. Analytics and lead-capture integration. Real codebase and deploy pipeline, not slideware.
Stack. React, TypeScript, Tailwind, deployed on Vercel with push-to-deploy from GitHub. I built it, deployed it, instrumented it, and kept it running. I am glad to walk through the code and the decisions live.
Every push to the repository deploys the site and submits the changed pages to search engines automatically. The ten most recent runs:
Systems that run behind the scenes
The same approach works inside a business: find the repetitive review, the scattered data, or the recurring handoff, then build a system that gives people a better starting point.
The problem. A manufacturer compared incoming CAD drawings against an approved set by hand. It was slow, easy to miss something, and a bottleneck when orders picked up.
What I built. A comparison workflow that flags mismatches so a person reviews the exceptions instead of every line. The judgment stays human. The grinding, line-by-line part does not.
Roughly 80% less time spent on drawing review, at the same accuracy.
The problem. Leadership believed on-time delivery was 68%. The metric was measuring the wrong thing, and the data was scattered across the system.
What I built. I cleaned and reorganized the workflow so it measured the right thing and stayed honest going forward. The fix was as much clear executive communication as data work.
The true on-time rate was about 90%, not 68%.
The problem. Staff dug through files to answer fit questions about load, span, environment, and use case.
What I built. An assistant that points them to the right product from structured requirements, turning scattered product knowledge into a repeatable answer.
The problem. Useful project information lived in too many places.
What I built. A layer that pulls the relevant details into the tools the team already uses, cutting the "where is that?" tax.
Where AI and automation fit
AI is one component, used where it earns its place and skipped where a plain rule is better.
Automated indexing and monitoring
The deploy-time indexing pipeline and the weekly scan run on their own and surface only what needs attention, instead of another dashboard nobody opens.
Research and content workflows
Finding the questions buyers actually repeat, shaping them into tool ideas, and turning messy research into shippable site copy and product concepts.
Structured scoring and triage
Scoring noisy inputs against a rubric, separating real signal from junk, and publishing a reviewable shortlist for a human to decide on.
An LLM handles scoring, summarizing, drafting, or classifying when it is the best fit. A deterministic rule handles the rest. What ships is the working system.
The work at a glance
| What | Problem | Built | Result |
|---|---|---|---|
| Client product site | No buyer resource or self-service for a manufacturer's catalog | 31-page site, 2 interactive tools, automated indexing pipeline, analytics and lead capture | Shipped to production, indexed, lead capture live |
| Drawing review automation | Manual CAD comparison bottleneck | Exception-based review workflow | ~80% less review time |
| Delivery-data repair | Leadership tracking a wrong on-time metric | Cleaned, reorganized delivery-data workflow | 68% to ~90%, corrected |
| Product advisor + configurators | Buyers and staff need quote-ready config | Guided in-product decision tools | Quote-ready output, buyer self-service |
| Product-fit assistant | Staff digging through files for fit answers | Guided recommendation assistant | Faster, repeatable fit answers |
| Indexing and monitoring | Slow indexing, no visibility | Deploy-time submission plus weekly scan | Automated, low-noise monitoring |
Methods I use
Build and ship
React, TypeScript, Tailwind, Node, and Python. Vercel deploys, Git and GitHub workflow, lightweight APIs and automation scripts, CSV and JSON data work.
AI and automation
LLM-assisted scoring, classification, and drafting. Prompt design and structured outputs. Human-in-the-loop review. Agent and tool workflows. Evaluation rubrics with pass/fail gates.
GTM and revenue ops
Lead-qualification and buyer self-service tools. Quote-prep flows. SEO and indexing automation. Google Search Console and GA4 workflows. Reporting into Sheets and Drive. Sales-handoff summaries.
Business translation
Sitting with nontechnical teams, mapping the workflow, finding the repeated decision, and shipping something they will actually use. Plain-English outputs. Industrial and field-service context.
Why this matters
Most teams do not need another AI demo. They need someone who can walk into a messy workflow, find the repeated decision, and build the smallest useful system around it.
That is the throughline here: buyers who need help choosing, sales teams tired of vague quote requests, operators comparing documents by hand, leaders working from a bad number, teams with useful data scattered everywhere.
The technology changes. The job stays the same: turn confusion into a better next step.
Want to see the work behind one of these?
I can walk through the logic, the workflow, the assumptions, and what I would improve next. The point is not that every tool is perfect. It is that they are real, specific, and built around actual business friction.