Brad Hobbs Dacula, GA

Websites & practical AI for small businesses

Make it easy for customers to reach you. Take the busywork off your plate.

I’m Brad Hobbs, a developer near Dacula, Georgia, and I’ve been building software for 15 years. I fix the parts of small-business websites that get in a customer’s way, and I set up small AI workflows for tasks your team repeats every week. Every project has a written scope and a handoff you can use.

Book a 15-minute call →
(912) 663-9223brad@bradhobbs.devReplies within one business day.
oakstreetplumbing.example
Concept homepage for a fictional plumber with a call button, towns served, and a quote form
The same concept on a phone, with a sticky Call now button
Concept · fictional business

Oak Street Plumbing concept

A website-refresh concept for a fictional local plumber, with tap-to-call, the service area on the first screen, and a short quote form.

1 / 8

Oak Street Plumbing concept

1 / 2
Concept homepage for a fictional plumber with a call button, towns served, and a quote form

On a computer

A website-refresh concept for a fictional local plumber, with tap-to-call, the service area on the first screen, and a short quote form.

Storefront door configurator with 3D preview

1 / 10
Full-screen 3D preview of a dark bronze double storefront door with a panic bar and a mail slot, seen at an angle

3D preview

Any configuration can be opened full screen and turned around in 3D. The model is built from the same dimensions the drawings use.

  1. Options show up on the model. This is the panic bar added on the Add-Ons step.
  2. The mail slot sits on the active leaf, in the bottom third, where it was placed.
  3. Switch between the dimensioned 2D drawing and the 3D model. The 3D view is built with Three.js through React Three Fiber.

Sentinel: a Windows trading workstation

1 / 7
One synthetic candlestick chart with a level table below and a watchlist beside it

Chart and level dock

The browser preview runs on deterministic synthetic bars and levels. The dock and watchlist keep related market context beside the chart.

  1. The browser capture is explicitly marked synthetic. These prices and levels are generated locally.
  2. The level dock lists status, price, distance from spot, volume, and source details.
  3. The watchlist stays available next to the active chart.

Community resource directory

1 / 2
Resource directory with a search box, category and city filters, and four result cards

On a computer

Searchable listings with category and city filters, a visible result count, and accessible form labels.

Pickem: live college football picks on every platform

1 / 9
Pickem live game view with the score, quarter and clock, possession, and a live-updates status bar

Live game view

Scores, the game clock, and possession update on their own while you watch. No refresh, no reload.

  1. While a game is in progress the app checks for new data every 15 seconds (every 60 seconds in the hours before kickoff), and only re-renders the games that changed.
  2. These screenshots use the app’s built-in test scenarios, so this banner marks the data as simulated. Real games use the same screen.
  3. Score, quarter, and clock come from a sports-data provider. A scheduled AWS Lambda pulls them every minute and stores them in the database.
  4. Possession and down-and-distance come from the same live feed.

Desktop + web shared views

1 / 1
Workspace notes app with search, a category filter, and three note cards

Desktop + web shared views

One set of React views shared by an Electron desktop app and its browser version.

AI-ready React workspace

1 / 1
Work-item queue with totals, filters, and three fictional items

AI-ready React workspace

A repository set up so AI-assisted changes arrive with specs, tests, and docs a person can review.

Workout tracker with an AI coach

1 / 9
AI coach chat answering which machine to use for the lower back and how to set it up

AI coach: gym questions

A real answer from the live app. The coach knows the gym’s actual machines, so it names one instead of giving generic advice.

  1. Each question is sent with the gym’s equipment list, the user’s plan, and this week’s workouts, so answers refer to real machines by model number.
  2. The chat goes through the app’s own AWS Lambda. The AI provider’s API key stays on the server and never reaches the browser.
Serving
  • Dacula
  • Hamilton Mill
  • Auburn
  • Buford
  • Braselton
  • Hoschton
  • Lawrenceville
  • Snellville
  • Gwinnett, Barrow & Jackson counties
AI workflow projects can be fully remote

Technologies

Search what I’ve worked with.

Type a tool, language, or platform, or tap a tag, to see the projects where I used it. Everything listed comes from a project in the Work section.

Popular

54 technologies across 11 projects

Frontend & mobile

Backend & APIs

AWS

Cloud & DevOps

Data & messaging

AI

Practices

Who I help

Local businesses where a clearer site or one less chore actually matters.

I work with owners who can point to a real problem, like calls that never come in, questions staff answer ten times a week, or a site nobody can update. If a fresh coat of paint won’t fix it, I’ll tell you.

i.

Home & local services

Plumbers, HVAC, landscapers, cleaners. People search on a phone and want to know if you serve their area and how to reach you.

“Do you come out to Auburn?”

ii.

Appointment businesses

Salons, clinics, studios, tutors. Customers need current hours, services, and a booking path that works.

“Are you open Saturday?”

iii.

Small professional firms

Accountants, attorneys, consultants. Visitors should be able to tell in seconds whom you serve and how to start.

“Is this firm right for me?”

What a refresh looks like

Same business. Fewer steps to call.

This is a concept built for a fictional plumber to show my approach. I start with what a customer is trying to do on a phone, then remove whatever gets in the way.

BeforeA typical older site: a generic welcome, no service area, and a phone number you can’t tap.
oakstreetplumbing.example
Redesigned concept homepage with a call button, towns served, and a quote form
Redesigned concept on a phone with a sticky call bar
After ConceptServices, towns served, and a one-tap call button are all on the first screen. View the concept site
  1. Say what you doA headline that names the services, not “Welcome.”was: “Welcome to our website!”
  2. Show where you workTowns served on the first screen, so nobody has to guess.was: “our community”
  3. Make calling one tapA call button up top and a sticky call bar on phones.was: a plain-text number in the footer

Services

Three small, clearly scoped ways to start.

Each one ends with something working and a written guide. We agree on what’s included, and what isn’t, before any work begins.

Website

Website refresh

A focused rebuild of the pages that matter, designed for phones first, so visitors understand what you do and can get in touch.

  • Up to five core pages
  • Mobile layout and clear contact path
  • Basic search setup (titles, descriptions)
  • Two rounds of feedback
  • Launch and a short update guide

Not included: online stores, copywriting, photography, paid ads, hosting fees. Quoted separately if needed.

Details: website refresh →
AI · Start here

Workflow review

One conversation about a task your team repeats. You get a clear picture of how it works today and what’s worth changing.

  • 60–90 minute interview
  • Simple map of the current workflow
  • Prioritized recommendations
  • One-page action plan
  • No tool purchase required

Good fit: inquiry replies, internal note summaries, organizing non-sensitive leads, routine first drafts.

Details: workflow review →
AI

One-workflow pilot

We choose one task, set it up in tools you approve, and test it on realistic examples before your team relies on it.

  • Setup in client-approved tools
  • Test cases, including the awkward ones
  • A person reviews anything customers see
  • Staff walkthrough
  • Written instructions

Measured, not promised: we record a baseline first and report what actually changed.

Details: one-workflow pilot →

Every project is fixed-scope and quoted in writing before any work starts. Start with a 15-minute call →

What an AI pilot looks like

AI does the first draft. You make the call.

Most small-business AI wins are ordinary tasks: replying to common questions, summarizing notes, sorting leads. I set up one of those at a time, in tools you approve, and nothing reaches a customer without a person checking it.

  • Uses your information. Drafts come from your service list, hours, and FAQ.
  • Stays in your hands. Approve, edit, or discard every draft.
  • Tested before launch. At least five realistic examples, including a bad one.
  • Measured honestly. We compare against how long it takes today.

AI assistant demo

See how a site assistant works, and how I’d protect it.

The chat button in the corner is a demo of the kind of assistant I build for business websites. It answers from the site’s own content and shows its sources. Here the answers are pre-written, so there are no AI costs and nothing to abuse. On your site, I’d connect it to a real AI model with the protections below.

Try it

Pick a question to open the assistant, or use the chat button at the bottom right.

Protections in a live version

A public chatbot needs limits, or bots and bored visitors will run up your bill. This is what I set up:

Bot check
Cloudflare Turnstile, verified on the server
Rate limits
Per visitor and per IP, e.g. 10 questions a day
Burst throttling
API Gateway request limits
Input limit
500 characters per question
Spending cap
Pauses the assistant at a monthly limit
Topic rules
Declines off-topic and “ignore your instructions” requests
Customer data
Not stored. Usage is logged without personal details.

This demo runs entirely in your browser. There’s no AI model behind it, and nothing you type is sent or saved.

How a live version answers a question

  1. Bot checkAn invisible Turnstile challenge, verified on the server.
  2. Rate limitPer-visitor and per-IP limits in DynamoDB, plus API Gateway burst throttling.
  3. Budget + input checksLength limits and a monthly spending cap that pauses the assistant.
  4. RetrieveFinds the closest sections of the site’s content using embeddings.
  5. AnswerClaude writes a short answer using only those sections, with citations.
  6. Check + logDeclines off-topic or “ignore your instructions” requests, and logs usage without personal data.

Steps marked with a diamond are = protection layer. It runs on AWS: API Gateway, Lambda, and DynamoDB.

Want one for your business? The same setup can answer questions from your service pages, policies, or internal docs, with the same limits and cost controls. It usually starts with a workflow review.

Talk about a site assistant →

How it works

No surprises. Everything written down.

You’ll always know what you’re getting, when, and how we’ll tell whether it worked.

  1. Talk

    A 15-minute call about what customers should do on your site, or which task eats your team’s time.

  2. Scope

    A written proposal: deliverables, exclusions, timeline, cost, and how we’ll measure success.

  3. Build & review

    I build in accounts you own and show you an early version before anything is final.

  4. Hand off

    Links, forms, and test cases are checked, then it launches with a short guide.

Work

Things I’ve built

Client work, employer work, personal projects, and portfolio samples, each labeled for what it is. Client and employer work is shared with written permission; screenshots use fake data and leave out private details.

Employer work · shared with permission
meeting-room app · session view
Meeting-room desktop app showing a session with its moderator and presenters, each with a Start button

Desktop + Web · Distributed systems

Event presentation platform: cloud, venue, and every room

Problem
At a large conference, hundreds of presentations have to reach the right room on time, often over hotel internet that drops out. Technicians need to watch and control every room at once, and speakers need an easy way to hand over very large files.
Approach
A cloud platform plus a small Kubernetes server at each venue that holds its own copy of the event and keeps running offline. RabbitMQ keeps the two databases in step, files are mirrored to on-site storage, and every meeting room runs a locked-down desktop app built from the same React code as the web apps.
What I built
I designed the architecture and led the team that built it, and I wrote major parts myself, including the meeting-room desktop app, the shared-views monorepo, and a developer tool that runs the whole stack locally. The platform includes a speaker upload app, an admin app, a kiosk-mode room app that runs PowerPoint, Keynote, PDF, and video, a live ops dashboard, .NET, Node, and Go services, and a virus-scanning file pipeline.

Built as an employee and shared with written permission. The code is private. The product name and logo are changed, screenshots use generated fake data from a local test environment, and the diagrams are redrawn.

Event presentation platform: cloud, venue, and every room

1 / 10
Meeting-room desktop app asking for the event PIN on a numeric keypad

Event PIN sign-in

Room PCs are shared, so technicians unlock them with the event’s PIN instead of a personal login.

  1. Connect to the on-site server, reach it remotely, or go straight to the cloud.
  2. The server checks the PIN and starts a session for that event only.
Client work · shared with permission
configurator · window wall in 3D
Configurator dimensions step for a custom window wall, with a 3D preview of four glass panels, a door, and a transom

Desktop · Electron + AWS

Storefront door configurator with 3D preview

Problem
Quoting a commercial aluminum door means turning an opening size into a frame, door leaves, glass, and hardware, then pricing all of it from a price list. Doing that by hand is slow, easy to get wrong, and the shop still has to work out the cut sizes again.
Approach
One desktop app built around shared pricing and manufacturing packages, so the estimator, the PDF, and the server all work from the same numbers. A step-by-step wizard, live 2D drawings and a 3D preview, and a small AWS backend for quotes, sign-in, and the price book, with a local cache so saves aren’t lost when the network drops.
What I built
I designed and built it on my own: the Electron app, the Fastify API, the MySQL database, and the AWS infrastructure in CDK. It covers 36 stock products (28 doors, 7 windows, and a custom window wall), dimensions from rough opening, frame, or door size, finishes, glass, and hardware, a 3D preview, porch layouts, customer and shop PDFs, emailed quotes with a customer approval link, a six-step order lifecycle, a price matrix, and an admin price book with an audit trail. Custom (non-stock) pricing is only partly mapped: some glass options still fall back to catalog prices, and only quote saves work offline.

Built for a client and shared with written permission. The code is private. The product name, logo, prices, and data are changed, the screenshots use fake data from a local test setup, and the diagrams are redrawn.

Storefront door configurator with 3D preview

1 / 10
Full-screen 3D preview of a dark bronze double storefront door with a panic bar and a mail slot, seen at an angle

3D preview

Any configuration can be opened full screen and turned around in 3D. The model is built from the same dimensions the drawings use.

  1. Options show up on the model. This is the panic bar added on the Add-Ons step.
  2. The mail slot sits on the active leaf, in the bottom third, where it was placed.
  3. Switch between the dimensioned 2D drawing and the 3D model. The 3D view is built with Three.js through React Three Fiber.
Client work · shared with permission
Sentinel · local synthetic preview
Sentinel chart workspace with a synthetic SPY candlestick chart, level dock, and watchlist

Desktop · Charting and supervised trading

Sentinel: a Windows trading workstation

Problem
Traders need to read charts, market levels, account state, and order checks together before making a decision. Those pieces are easy to lose when they live in separate tools and data feeds.
Approach
I built a Tauri desktop shell with React workspaces, a TypeScript gateway that normalizes market data, and a Go service that owns trading state and risk checks. Native access and broker credentials stay outside the web view.
What I built
I designed and built Sentinel solo. The code includes configurable chart layouts, studies and drawings, a level dock and watchlist, dashboard panels, an options chain and portfolio view, and an order ticket that leads to a separate review. The Go engine checks fresh broker evidence and coded limits before review and again before submission. Native market and broker features require their configured services; these captures use synthetic or illustrative development data.

The code is private and this client work is shared with written permission. Chart images use Sentinel's deterministic browser data; options images use its development-only illustrative fixtures and a fake masked account. No live broker order was placed.

Sentinel: a Windows trading workstation

1 / 7
One synthetic candlestick chart with a level table below and a watchlist beside it

Chart and level dock

The browser preview runs on deterministic synthetic bars and levels. The dock and watchlist keep related market context beside the chart.

  1. The browser capture is explicitly marked synthetic. These prices and levels are generated locally.
  2. The level dock lists status, price, distance from spot, volume, and source details.
  3. The watchlist stays available next to the active chart.
Sample · fictional data
community-resource-directory
Resource directory with a search box, category and city filters, and four result cards
The directory on a phone with stacked filters

Web · Accessibility

Community resource directory

Problem
People looking for local help need to filter quickly by need and city, and see right away how many options match.
Approach
Keep the public side small and clear, with one search box, two filters, a visible result count, and proper labels and focus states.
What I built
A React and TypeScript directory with automated tests. The records are fictional, and it hasn’t had a formal accessibility audit.

Community resource directory

1 / 2
Resource directory with a search box, category and city filters, and four result cards

On a computer

Searchable listings with category and city filters, a visible result count, and accessible form labels.

● Personal project · private beta
pickem · live game view
Pickem live game view showing Alabama at Georgia in the third quarter, with a live-updates status bar
Pickem on a phone with Georgia picked against the spread for 50 points

Mobile + Web · AWS

Pickem: live college football picks on every platform

Problem
A pick’em league needs picks, point totals, lock times, and live scores in one place, on whatever phone or computer each player uses, and the rules have to hold up on game day.
Approach
One Expo React Native codebase for all three platforms, on a serverless AWS backend. Lock rules are enforced on the server, and live scores flow from a sports-data provider to every open screen without a refresh.
What I built
Players pick games against the spread and wager points from a weekly budget. Game cards go live on their own: a scheduled AWS Lambda pulls scores every minute, and the app checks for changes every 15 seconds during games. Also: a standings table, graded history, a commissioner slate builder, Google and Apple sign-in, and biometric unlock on phones.

The production code is private. The linked repo is a small, rewritten sample of the shared-codebase pattern, with no backend, accounts, or live data.

Pickem: live college football picks on every platform

1 / 9
Pickem live game view with the score, quarter and clock, possession, and a live-updates status bar

Live game view

Scores, the game clock, and possession update on their own while you watch. No refresh, no reload.

  1. While a game is in progress the app checks for new data every 15 seconds (every 60 seconds in the hours before kickoff), and only re-renders the games that changed.
  2. These screenshots use the app’s built-in test scenarios, so this banner marks the data as simulated. Real games use the same screen.
  3. Score, quarter, and clock come from a sports-data provider. A scheduled AWS Lambda pulls them every minute and stores them in the database.
  4. Possession and down-and-distance come from the same live feed.
● Personal project · live
The workout tracker’s AI coach turning a request into a 30-minute cardio plan card

Web app · AI · AWS

Workout tracker with an AI coach

Problem
A flexible weekly routine didn’t fit fixed-schedule apps, none of them offered useful recovery guidance, and generic AI chat didn’t know my plan or my gym’s machines.
Approach
A calendar-first log, a rules engine that reads each workout and recommends recovery with its reasons, and an AI coach that gets the plan, this week’s workouts, and the gym’s equipment list with every question.
What I built
A multi-user React app with per-set logging, a recovery coach, a stretch guide matched to the muscles worked, and 8-week progress charts. The AI coach runs through the app’s own AWS Lambda, so API keys never reach the browser. It answers in plain language or returns a complete workout in the app’s own format, ready to add to today. It can use Claude or OpenAI; the live app currently uses OpenAI’s GPT-4o mini. Hosted on Lambda, API Gateway, S3, and CloudFront, defined with AWS CDK.

Workout tracker with an AI coach

1 / 9
AI coach chat answering which machine to use for the lower back and how to set it up

AI coach: gym questions

A real answer from the live app. The coach knows the gym’s actual machines, so it names one instead of giving generic advice.

  1. Each question is sent with the gym’s equipment list, the user’s plan, and this week’s workouts, so answers refer to real machines by model number.
  2. The chat goes through the app’s own AWS Lambda. The AI provider’s API key stays on the server and never reaches the browser.

See all projects →

Brad · Dacula, GA

About

Hi, I’m Brad. I write it all down.

I’ve been building software for 15 years, including web and desktop apps, mobile apps, cloud systems on AWS, and tools that help teams work faster with AI. Now I’m bringing that experience to small businesses around Dacula, Hamilton Mill, Buford, and the rest of Gwinnett. I do the work myself, so the person you talk to is the person building your project.

  • You own everythingYour domain, content, and accounts stay in your name.
  • No inflated promisesNo guaranteed rankings or revenue. We measure and report honestly.
  • Your data stays yoursNothing goes into an AI tool without your written OK.
  • A human in the loopA person approves anything a customer will see.

More about how I work →

FAQ

Common questions

Do I need to know what I want before we talk?

No. Tell me what’s frustrating, like missed calls or repeated questions. I’ll suggest the smallest project that could help, or tell you if I’m not the right fit.

Will you write the text for my site?

You supply the facts: services, hours, service area, photos. I organize and edit them for the web. Full copywriting is quoted separately.

Is AI safe to use with my customers’ information?

We decide together what data a tool may see, and sensitive information stays out unless there’s a proper plan for it. A person reviews anything customers will see.

Who owns the website when we’re done?

You do. The domain, hosting, and content are in your name, and you get a guide for making simple updates.

How long does a website refresh take?

It depends on how quickly content and feedback come in. The timeline is written into your scope before we start.

Can you guarantee more leads?

No one honestly can. We agree on what to measure, record a starting point, and I report what changed.

Next step

Tell me what’s getting in the way.

Pick whichever is easiest. If there’s a good fit, I’ll send a short written scope. If not, I’ll tell you, and you’ll leave with at least one useful idea.

I reply within one business day.