Back to work
AI evaluation hackathon
AI Response Evaluation Hackathon
A practical evaluation flow for comparing AI responses with rubrics, scoring and review-friendly structure.
Challenge
- AI responses can sound confident while still being incomplete, biased or wrong.
- Reviewers need a repeatable way to compare outputs instead of relying on gut feel.
- Prompt iteration is slow when quality signals are scattered across notes, spreadsheets and ad hoc comments.
Solution
- Built the concept around explicit rubrics for correctness, usefulness, bias and clarity.
- Separated scoring, review and iteration so each response can be evaluated with consistent criteria.
- Used lightweight automation thinking to turn evaluation from a one-time hackathon task into a reusable workflow.
Outcome
- Prompt changes become easier to compare because the scoring model gives teams a common language.
- Weak responses are easier to diagnose, not just reject.
- The project shows how AI can be made practical by wrapping it in human-readable evaluation systems.
Stack and patterns
Built as an evaluation pattern that can plug into future AI-assisted workflows.
PythonLLM promptsEvaluation rubricsScoring modelReview dashboard conceptsIteration logs
AI output is useful but hard to trust consistently.
People can see potential, but nobody has a shared way to judge quality, bias, usefulness or failure modes.
Define the rubric before automating the workflow.
Start with what good means, how reviewers score it and where the feedback should improve the next prompt.
Real prompts, outputs and review disagreements.
A strong evaluation flow needs examples that reveal where judgment currently gets fuzzy.
Previous proofFactory Layout / React Leaflet WidgetFactory planning concept / 2026
A planning surface that turns static layout knowledge into something teams can explore.
Next proofSlot Machine Game ExperimentFrontend experiment / 2026A compact interaction lab for testing timing, state and rewarding feedback loops.
My process
A clear path from idea to impact.
Clarify
We define the problem, users and constraints before anything gets overbuilt.
Prototype
I design the core workflow quickly so direction becomes visible early.
Build
I build clean, scalable foundations with attention to detail.
Ship
We launch, test with real users and make sure it actually works.
Improve
I iterate, optimize and evolve the product with your team.
Practical. Collaborative. Built to last.
Let's work together
Tell me what you want to build.
Let's build something useful.Project brief
Shape the first message.
Choose the build, name the friction, then send Pontus a brief that already has structure.Hi Pontus, I am interested in: A conversion-focused website. Main friction: The current experience does not create enough confidence. Timeline: Soon. Context: What makes the current site, tool or workflow feel less trustworthy than it should? Useful starting point: Pontus should review the current offer, page structure, audience and strongest proof points. Best,
What Pontus will ask next
- Who needs to trust you faster?
- What should visitors do next?
- What proof already exists?
A conversion-focused websiteSoon