Piotr Mrozinski
Let’s Talk
Back to portfolio

I Don’t Just Sell Technology. I Build With It.

I wanted a better way to run my job search. So I built one.

What started as a personal problem became an AI-powered Career Agent.

I didn't want another tool that simply matched keywords or asked a model to rewrite my CV.

I wanted a system that could discover relevant roles, understand what actually made them a fit, use verified evidence from my career, generate tailored applications, challenge its own output, repair unsupported claims and track what happened next.

So I built it.

The System
  1. 01Discovery
  2. 02Fit Engine
  3. 03Career Knowledge Base
  4. 04Evidence Guard
  5. 05CV Writer
  6. 06Multi-model Evaluation
  7. 07Factual Verifier
  8. 08Repair Loop
  9. 09Final Artifact
  10. 10Application Tracking
Model providersCursor SDK · OpenAI · xAI/Grok · DeepSeek
Production systemAPI · SQLite · Scheduled workflows · Persistent run artifacts
Engineering disciplineGit · Automated tests · Model routing · Bounded repair

Different parts of the system do different jobs.

Models can be swapped and routed depending on the task. Career evidence is separated from generation. Outputs are checked against a controlled knowledge base. Failed verification can trigger repair rather than simply accepting plausible-looking output.

The result is a working system with an API, persistent database, scheduled workflows, automated tests, model routing and durable run artifacts.

Not a prompt. A system.

Build by the Numbers

35,635Lines of source code
139Source files
660 / 660Automated tests passing
4 model providersCursor SDK · OpenAI · xAI/Grok · DeepSeek
9 pipeline stagesFrom discovery and fit through evidence, generation, verification, repair and final artifact
86 commits · 10 active daysFrom first commit to the current production system
1 actual userMe. Every day. 😄

Why I Built It This Way

AI can generate impressive-looking answers very quickly.

That doesn't mean they're right.

For something like a CV, a plausible invented claim is worse than a mediocre sentence. So I designed the system around evidence, verification and controlled generation, rather than trusting a single model to get everything right.

The same principle applies to how I think about enterprise technology:

Understand what the system is actually doing, where it can fail, and what controls make it useful in the real world.

Under the Hood

Multi-model architectureCursor SDK · OpenAI · xAI/Grok · DeepSeek
Evidence-controlled generationCareer Knowledge Base · Evidence Guard · Factual Verifier · Repair Loop
Production systemAPI · SQLite · Scheduled workflows · Persistent run artifacts
Engineering disciplineGit · Automated tests · Model routing · Bounded repair

Built with AI. Designed with evidence controls. Tested like software. Used in production.

Why This Matters to How I Sell

I'm not trying to become a software engineer.

I'm becoming a salesperson who can operate much closer to the technology.

I can talk to a founder, engineer or technical buyer without needing every concept translated into sales language first — and then connect that technology back to the commercial problem the customer is trying to solve.

The technical founder shouldn't have to dumb the product down before the salesperson can sell it.