OO.

SOFTWARE · DATA · APPLIED AI

Projects across
software & AI.

A collection of work in software, data, and applied AI—focused on practical problems and systems people can inspect and use.

Browse projects Find me on GitHub
PROJECT FLOW / 001PROBLEM → BUILD → REVIEW
01 / DEFINEUnderstand itProblem & context
02 / BUILDTry an approachTools & methods
03 / REVIEWCheck the resultTests & evaluation
SOFTWARE✳DATA · AI
PORTFOLIO / 2026INDEPENDENT PROJECTS & EXPERIMENTS

01 — SELECTED PROJECTS

Useful systems,
made inspectable.

Two hands-on projects exploring practical AI workflows, with clear evidence of what each system does.

FEATURED PROJECTGENAI / MLOPS

Company Knowledge
Assistant

Ask a question in natural language. The assistant can retrieve document passages, query a read-only analytics dataset, or use focused tools. I built it to explore these pieces and make the results inspectable.

Hybrid RAGTool callingEvaluationObservability
Open live demo More on GitHub

NOTE BrightBlocks Toys is fictional. Its documents and retail figures are synthetic demo data.

QQUESTION01

Which product led net sales in 2025?

AGROUNDED ANSWER02

Harbor Rescue Kit led with €79,066.80.

↳ Read-only SQL query · 3 rows returned
↳ Synthetic 2025 retail dataset
RETRIEVAL EXPERIMENTHybrid search
1.00 hit@5

On a small synthetic evaluation set. A useful baseline, not a claim of production-scale performance.

01

Two kinds of knowledge

Document retrieval handles policies, supplier records, product information, and quality reports. A separate read-only SQL tool answers questions over synthetic sales data.

02

Evidence is visible

Answers include citations or a record of the tool used, so a person can inspect how the assistant reached its response.

03

Quality is measured

Retrieval strategies can be compared against labeled questions. Application tests run in CI before deployment; MLflow records evaluation runs.

PROJECT 02CLASSIFICATION / AGENT TOOLS

Routewise
Support Triage

A customer message is matched to a likely issue, sent to one of six fictional support teams, and turned into a useful next step. Requests the router cannot confidently place are held for human review.

Text classificationHuman reviewTool useFastAPI
Try the live demo How it works

DEMO Uses BANKING77 by PolyAI (CC BY 4.0). Teams, accounts, and card actions are fictional; no real banking systems are connected.

ROUTING EXAMPLE01 / 03
CUSTOMER MESSAGE

“My card doesn’t work when I try to pay.”

LIKELY ISSUECard payment problem
SEND TO THIS TEAMCards & Wallets

Found a help article and prepared guidance.

OPTIONAL DEMO Lost-card requests can try a confirmed simulated freeze.
OVERALL HELD-OUT TEST SETCorrect specialist
96.4%

Separately, a review threshold set to target 98% accuracy on validation data routed 93.4% of test requests.

01

Choose a team—or review

A supervised classifier maps messages to 77 support topics, then to six specialist teams. A review threshold avoids forcing a team choice on uncertain requests.

02

Show the specialist’s work

The selected team searches source-linked help guidance and prepares a next step. Visitors can inspect the activity trace and compare a simple tool-matching baseline.

03

Make actions safe to explore

For a lost-card request, visitors can confirm a simulated freeze. It changes only temporary demo state, with a visible result and audit entry.

02 — FOCUS

Useful work,
clearly explained.

I’m interested in projects that connect a clear user need with thoughtful implementation, useful data, and results that can be checked.

01
Start with the problem

Understand the task and context before choosing a technical approach.

02
Make behavior measurable

Use tests and evaluation to see where a system works and where it needs improvement.

03
Explain the choices

Make the design, sources, and tradeoffs understandable to the people using the result.

TOOLS ACROSS PROJECTS

Selected technologies used in this portfolio.

PythonFastAPIscikit-learnReactPostgreSQLChromaMLflowDockerGitHub Actions