How I Built Two Production AI Systems in Two Weeks (While Working Full-Time) … and why it matters

Let me tell you a real story of pain and sorrow, and how I solved it fast with technology. So fast that it changed my world view. Maybe it will change yours as well.

Picture this: I live in Egypt and the time had come for my Polish parents to visit me. They don’t speak English. My wife is Egyptian and doesn’t speak Polish. Guess who for 2 weeks was the only connection of my parents to their new reality? Every interaction that they wanted to make had to be made by me. And oh boy, being a son that wants to host parents is difficult enough, but adding to this being a personal translator changes things to an extreme difficulty level. Imagine a political discussion about Arab immigrants in Poland between my dad and Egyptians. Things were getting heated while in between that was yours truly, Paweł, translating everything word by word. In many cases not agreeing with what needed to be said. If I wanted to say something, I needed to say it in two languages at the same time because two sides wanted to participate. Both sides of course were responding to my sentence and all of a sudden we have two separate threads of conversation. This didn’t work and caused me a lot of frustration. Best option was for me to be just a translator. Period.

This got me thinking and in the spur of the moment I decided to try to solve my problems with AI and technology … and it changed the way I perceive current technological transformation.

We have a family chat on WhatsApp where my parents, my wife and I share messages. Of course they speak Polish, Fatma English and me sometimes PL and sometimes EN language. As you can imagine it is not the most flowing chat in the world. So I made a decision. I’ll write using AI a translation bot that will join the chat and whenever message comes in PL translates to EN, whenever EN translates to PL. To my surprise half a day later I had deployed a solution that was tested and costs close to nothing.

This got my attention. So I decided to go one step further - how about solving the translation problem so that my family could really communicate? Live translation.

It took me 7 next evenings (working full time, learning to play 2 instruments and spending time with friends in addition) to build a working solution and deploy it to Production. Fully tested. With failovers. I have a web application that does live speech-to-text and translation, has implemented 4 TTS providers (text-to-speech). Because some have Egyptian better, some English, some other and it would be nice to have failover … and because it was so easy that why not? Three translation providers and translation routing based on languages. For instance PL -> EN will go to provider A (as he does it the best) and PL to AR will go to provider B (also with failover). Of course detailed cost management, easy invitation system and everything with full test code coverage.

Yesterday my dad watched ping pong match in English with my application doing live speech to text, then translation to Polish and then text to speech. Ah, did I mention that I added separate STT (speech to text) in one afternoon? And guess what, also fully tested.

Total development time: Less than two weeks of evening work. While also: Cost: ~$100/month for Claude Code and ongoing API costs for STT, TTS, translation.

Here’s exactly how I did it—and how you can build similar solutions for your business problems and for yourself.

Quality metrics across both systems:

• 689 automated tests
• 98.5% passing rate
• 100% coverage on critical paths (unit, integration, E2E tests)

Compare this to traditional development: months of work, rigid scripts, breaks when anything changes.

Table of Contents

• Why This Matters for Your Business
• What Makes This Different: The Agentic Architecture
• My Agent Team: The Actual Configurations
• How This Applies to Your Business Problems
• What You Can Build Right Now
• The Real Cost (And What I Actually Paid For)
• It’s David and Goliath Time
• The Single Most Important Realization
• What Happens Next

Why This Matters for Your Business and you

If you’re in marketing, operations, customer service, or sales or as a son/daughter/husband/father custom applications and workflows are available for you for a fraction of the cost and time that they have been half a year ago. I can say with confidence that whatever business process you have or any business problem it can be solved cheap and fast.

And quality? I’m a software engineer and probably I wouldn’t write this application better.

• Marketing teams manually reformatting campaign data across platforms
• Operations managers copying information between CRM, email, and spreadsheets
• Customer service handling repetitive questions that follow predictable patterns
• Sales teams updating multiple systems after every client interaction

These aren’t “technical problems requiring developers.” They’re workflow problems where AI agents excel.

The old approach: hire developers, wait months, spend thousands, get something rigid that breaks when requirements change.

The new approach: describe what you need, get one technical savvy person that knows how to use AI and you will get that for a fraction of the cost. I believe that we are witnessing birth of new role in organization - AI automation expert. He/she will write you a custom application, webpage, automate your marketing processes or integrate systems that previously you could only dream to connect.

What Makes This Different: The Agentic Architecture

Imagine a system where you have 5 specialized AI systems and each prompt to do one thing very good.

One can be Project Manager, one architect, one software developer, one QA engineer, one DevOps. If that’s got too technical too fast for you don’t worry. I’ll explain also outside of IT domain.

Where is the difference between normal AI prompting?

Well your Project Manager can orchestrate between different flows. Know when you ask for a new feature and utilize totally different flow to orchestrate AI agents working in parallel to complete your tasks. He will know that if this is a bug request flow will be different.

With marketing and blog posts? Imagine that you want to write a blog and you create 3 agents. One to help you with the blog writing, one specializing in finding best hook so your readers will keep engaged and a critic that will check your company guidelines and in general best practices. Your job is to write a story that doesn’t need to be grammatically sound. Your agents will fix that for you and deploy once they polish your article. You still need to provide them content as no one wants to read AI generated fluff but believe me that you would be surprised how easy it is and good.

Agentic AI uses specialized agents that understand context, adapt to changes, and coordinate autonomously—systems where specialized AI agents collaborate on complex tasks without human intervention. Also with current system limitations of context window (how much AI can hold in working session) every agent will have their own separate one. Saving you trouble of opening few chats for one task.

Ah, did I mention that I didn’t write those agents myself? Of course I asked AI to help me :)

My Agent Team: The Actual Configurations

Here are the exact agent configurations I use for both application development and content creation. These aren’t hypothetical—they’re the actual working agents that built the systems described in this post. Only thing that I didn’t put here is my system information as you wouldn’t benefit from that at all :)

For Application Development (7 Specialized Agents)

Project Manager

• Role: Orchestrates planning and execution from concept to deployment
• Tools: Read, Write, Bash, Grep, Glob, Task (delegates to other agents)
• Responsibilities:
  • Clarifies requirements and coordinates team
  • Creates plans and tracks progress
  • Ensures code reviewed, tests passing, deployed, docs updated
  • Delegates to specialists: BA, Architect, Developer, QA, DevOps, Writer
• Success criteria: Task complete when user accepts deliverable + all tests pass + code deployed + docs updated

Business Analyst

• Role: Defines requirements, user stories, and acceptance criteria
• Tools: Read, Write, Grep, Glob
• Responsibilities:
  • Analyzes user goals and identifies target users
  • Writes user stories with prioritization (MoSCoW method)
  • Creates comprehensive test cases covering happy path, errors, edge cases, security, performance
  • Validates feasibility with architect
• Deliverables: User stories + prioritized test case list + requirements document + user flows + edge cases + success metrics

Software Architect

• Role: Reviews architecture and designs technical solutions
• Tools: Read, Write, Grep, Glob, Task
• Responsibilities:
  • Provides architecture overview and design deliverables
  • Reviews code for patterns, security, scalability, testing coverage
  • Enforces standards: async/await, service layer, sequential migrations, proper error handling
  • Creates architecture diagrams, API specs, database migrations, technology justifications
• Focus areas: Security (no SQL injection, XSS, CSRF), scale (caching, connection pooling), patterns (router-service, provider abstraction)

Full-Stack Developer

• Role: Implements features with clean, tested code
• Tools: Read, Write, Edit, Bash, Grep, Glob, Task
• Responsibilities:
  • Follows TDD workflow (Red→Green→Refactor)
  • Implements backend (FastAPI + Pydantic + async/await + type hints)
  • Implements frontend (React hooks + PropTypes + i18n for all text)
  • Validates input, parameterizes queries, sanitizes output, checks auth
  • Requests architect review before QA handoff
• Testing targets: Critical 100%, Unit 95%+, Integration 90%+

Automation QA Engineer

• Role: Writes and runs automated tests at all levels
• Tools: Read, Write, Bash, Glob, Edit
• Responsibilities:
  • Writes unit tests (fast, no I/O), integration tests (Redis/DB), E2E tests (user journeys)
  • Runs all tests and reports failures with clear diagnostics
  • Debugs and fixes test issues while preserving intent
  • Ensures all tests pass before completion (98.5%+ pass rate)
• Test organization: Backend (unit tests in tests/unit/, integration in api/tests/), Frontend (component tests with Vitest + React Testing Library)

DevOps Engineer

• Role: Handles deployment, infrastructure, and verification
• Tools: Read, Write, Edit, Bash, Grep, Glob
• Responsibilities:
  • Rebuilds containers (always —no-cache for frontend)
  • Restarts services and verifies deployment thoroughly
  • Runs database migrations via psql
  • Performs mandatory pre-handoff verification: files exist + current timestamps + code changes confirmed + services responding + logs checked
• Critical: Always verifies deployment to catch cache issues early

Technical Writer

• Role: Maintains project and product documentation
• Tools: Read, Write, Grep, Glob
• Responsibilities:
  • Updates core documentation at task completion only
  • Maintains API endpoints, database schema, architecture docs, changelog
  • Keeps documentation minimal and maintainable
  • Deletes temporary tracking files (no PHASE_*.md, *_SUMMARY.md, *_REPORT.md)
• Philosophy: Focus on “what does it do?” not “how did we build it?”

For Blog Content (3 Specialized Agents)

Blog Writer

• Role: Generates effective, brand-consistent blog posts for business audiences
• Tools: Read, Write, Edit, Glob, Grep, WebSearch, WebFetch
• Voice: Clear, direct, neutral, human-readable plain English (multilingual-friendly)
• Content guidelines:
  • Begins with strong hook (question, fact, or bold statement)
  • Opens with problem/desire + solution preview in first 150 words
  • Organizes for easy scanning: short paragraphs, clear subheadings, bullets, numbered lists
  • Educates without overstatement—backs claims with logic or examples
  • Provides actionable steps and real-world examples
  • Maintains helpful, informative tone—no hype, fluff, or clichés
• Structure templates: Problem→Solution→How-To, numbered lists (“X Ways to…”)

Brutal Blog Critic

• Role: Rigorously analyzes and critiques blog content
• Tools: Read, Glob, Grep, WebSearch, WebFetch, Task
• Framework-driven approach:
  • Applies known structures (Problem→Solution→How-To, What→Why→How, AIDA)
  • Evaluates hook and headline: Does it grab attention? State problem? Promise solution?
  • Checks clarity and structure: Concise paragraphs? Descriptive subheadings? Short sentences?
  • Validates examples and value: Are claims supported? Tips actionable? Evidence present?
  • Assesses tone: Factual and helpful without hype?
• Feedback style: Brutally honest, detail-oriented, blunt but constructive with specific fixes

Hook Expert

• Role: Generates 3-5 compelling opening hooks for blog posts
• Tools: Read, Grep, WebSearch, WebFetch
• Hook strategies: Question, surprising statistic, mini-story, bold statement, contrarian angle, light-hearted quip
• Core principles:
  • Specificity over generality: Use concrete numbers, names, scenarios, vivid details
  • Credibility first: Ground every hook in reality with real statistics or plausible scenarios
  • Logical flow: Each hook must naturally lead into the blog topic
  • Tone consistency: Professional, efficient, human voice—helpful and specific
• Evaluation checklist: Relevant? Clear? Credible? Original? On-brand?

How These Agents Work Together

For application development, the Project Manager orchestrates the workflow:

1. Planning phase: PM → BA (requirements) → Architect (design review)
2. Implementation phase: PM → Full-Stack Developer (code) → Architect (code review)
3. Quality phase: PM → QA (tests) → DevOps (deploy + verify)
4. Documentation phase: PM → Technical Writer (update docs)

For blog content, the workflow is simpler:

1. I write rough draft with ideas
2. Blog Writer polishes structure, tone, and clarity
3. Hook Expert generates 3-5 compelling opening hooks
4. Brutal Blog Critic reviews everything with frameworks
5. Blog Writer refines based on critique
6. Repeat steps 4-5 until critic approves

Key insight: Agents handle execution (code, writing, testing, deployment). I provide domain expertise (business requirements, editorial judgment, strategic decisions, ideas). The bottleneck isn’t technology—it’s clearly articulating what I need.

How This Applies to Your Business Problems

Example: Customer Inquiry Workflow

Traditional approach:

1. Customer emails inquiry
2. Support reads email
3. Support checks CRM for customer history
4. Support searches knowledge base
5. Support drafts response
6. Manager approves response
7. Support sends email
8. Support logs interaction in CRM

Custom application / workflow that can be created in days (not weeks / months)

Input → Customer inquiry arrives

Processing:

• Monitor agent detects new email, extracts key details (product, issue type, urgency)
• Research agent pulls customer history from CRM, searches knowledge base for relevant articles
• Response agent drafts personalized reply using customer data and knowledge base content
• Routing agent decides: standard inquiry (auto-send) or complex issue (flag for human review)

Output → Response sent + CRM updated + ticket created if needed

Result: 80% of routine inquiries handled in under 2 minutes. Human agents focus on complex cases requiring judgment.

I don’t expect that this particular example will save you or your business. What I’m trying to do here is to open your mind for possibilities. What was available yesterday to big players with money today is available to you and me.

What You Can Build Right Now

Again this is just a set of ideas to get you thinking could I do it? Even if not by your own hands there are people willing to help you :)

For Marketing Teams

• Content repurposing system: Blog post → social posts → newsletter → LinkedIn article
• Campaign performance reporter: Pulls data from Google Ads, Meta, email platform → generates weekly summary
• A/B test analyzer: Reads results → calculates statistical significance → drafts recommendation email

For Operations Teams

• Onboarding coordinator: New hire confirmed → sends welcome email → creates accounts → schedules orientation → notifies team
• Vendor communication: Purchase order approved → generates PO document → emails vendor → logs in system → sets follow-up reminder
• Compliance checker: Scans contracts for required clauses → flags missing items → generates remediation list

For Customer Service Teams

• Inquiry router: Reads customer email → checks account history → categorizes issue type → routes to correct team → sends acknowledgment
• Knowledge base updater: Identifies repeated questions → drafts new help article → submits for review
• Follow-up scheduler: Resolved ticket → waits 3 days → sends satisfaction survey → logs response

For Sales Teams

• Meeting notes processor: Call recording → transcribes → extracts action items → updates CRM → schedules follow-ups
• Proposal generator: Sales call notes → pulls relevant case studies → generates customized proposal → formats PDF
• Pipeline cleaner: Stale opportunities → sends re-engagement email → moves to “Nurture” if no response

All of these are possible today. No development team required.

The Real Cost (And What I actually paid for)

Claude Code license: $100

LiveTranslator:

• Development: 7 evenings (~20 hours total)
• API costs: ~$1/month (speech-to-text, translation). Those are costs as I go. The more I will use the app the more I will pay.
• Hosting: $6/month (server + database)
• Total first month: ~$7

WhatsApp bot:

• Development: 1 weekend (~8 hours)
• API costs: ~$1/month (translation only)
• Hosting: Piggybacks on same server
• Total first month: ~$1
• Ongoing: ~$1/month

Combined monthly cost: ~$108 and about 30hrs of work for two production systems.

Traditional development estimate:

• Developer rate: $100-200/hour
• Project timeline: 2-3 months
• Backend developer: 120 hours @ $150 = $18,000
• Frontend developer: 80 hours @ $150 = $12,000
• Testing/deployment: 40 hours @ $150 = $6,000
• Total: $36,000 minimum
• Timeline: 3-6 months (if everything goes smoothly)

I hope that this starts to open a crack in your head and opens you to possibilities.

The savings aren’t just money. It’s time to market, ability to iterate quickly, and solving problems as they arise instead of waiting months for a scoped project.

It’s David and Goliath Time

AI powered David v Industry

Services that previously required large budgets are now accessible to small and medium businesses and passionate individuals. You can now get professional applications, websites, and automated internal processes at a fraction of traditional costs—often delivered faster than before.

This is your competitive moment. The gap between solo operators and well-funded competitors is narrowing. You can now:

• Build production-grade applications without a development team
• Maintain consistent branding across all channels
• Automate repetitive business processes
• Create professional content at scale

You still need to bring ideas, expertise, and judgment. But the technical barriers that used to limit small businesses are coming down.

The trend will only accelerate. The tools will get better, easier to use, and more accessible. The question isn’t whether small businesses will use AI agents—it’s whether yours will lead or follow.

Plus bonus note for all those enthusiasts and geeks. Those tools made technology fun and accessible for me. If it wouldn’t be for the AI I wouldn’t even dream of solving my day to day problems with technology. Today I have million ideas what to do next and what to build next. This is more than fun!

The Single Most Important Realization

You don’t need to understand how the AI works. You need to understand your problem clearly enough to explain it and either to an AI or a developer that works with AI.

That domain knowledge is more valuable than coding skills.

When I built LiveTranslator, I didn’t write the speech recognition code. I didn’t implement the translation algorithms. I didn’t configure WebSocket streaming.

I described what needed to happen. AI agents built it. Still this technology is at its infancy and technical background is necessary. System can make stupid mistakes, forget about features, make not the best decisions that’s why AI automation engineer can help you with that.

The bottleneck isn’t technology anymore. It’s clearly articulating what you need.

What Happens Next

My parents’ next visit is in two months. Dad will tell stories, my mother-in-law will share family news, and I’ll actually participate in the conversation instead of translating every sentence.

Our WhatsApp group now flows naturally across languages. Birthday planning, garden updates, late-night jokes—all happen without me as the middleman.

These aren’t tech demos. They’re production systems solving real problems.

You have similar friction points in your business. Manual processes eating hours every week. Information stuck in silos. Repetitive tasks that never reach the top of your team’s priority list.

Pick one. Describe it to Claude. Build your first agent system this week.

Start with something small and annoying. The marketing report that takes 45 minutes every Monday. The customer follow-up emails you keep forgetting to send. The data export you manually reformat every Friday.

Two weeks from now, that task could be automated.

The technology is ready. The tools are available. The only question is: which problem will you solve first?