Final-year Computer Science student at the University of Bucharest. I build data pipelines and tools, mostly in Python and JavaScript.

dra-solo.com · Email

My bachelor's thesis, still in progress. It collects macro data (FRED), market data (Yahoo Finance) and news into SQLite, flags anomalies with z-scores, and uses Claude to extract events and write hypotheses about which sectors a shock will hit. Each hypothesis has to quote its evidence from the input, and causal chains come from a fixed set of templates, not from the model.

Measured so far, on 35 hypotheses from three historical shocks (COVID 2020, the 2022 energy shock, SVB 2023):

• 0.246 of evidence spans are unsupported by the input, against 0.923 for an unconstrained LLM given the same signals.
• No fabricated numbers, sectors or causal chains.
• Moving one rule from the prompt into the output schema cut its violation rate from 91.4% to 0%.
• 271 tests run offline, and CI fails if grounding gets worse.

Next: retrieving similar past episodes to inform each hypothesis, an agent that tries to disprove hypotheses by checking real data, and an API. The code is private while I finish the thesis.

Python Claude API SQLite pandas Streamlit pytest

My own life tracker, still in progress. Habits and training are live, the food catalog is loaded but has no API yet, and spending and career tracking come next. One Cloudflare Worker serves both the site and its API from a D1 database. NFC tags around my home log a habit with one tap. Logs are append-only, and totals are computed when read, never stored. There is no build step and no npm dependency. GitHub Actions deploys it, and a smoke test checks every route.

JavaScript Cloudflare Workers D1 (SQLite)

Coursework built with OpenCV and scikit-learn. One assignment scores a Qwirkle game from a photo of the board: it straightens the board with a homography, classifies each tile and validates the moves. The other detects cartoon faces with HOG features, a linear SVM and a sliding window.

Python OpenCV scikit-learn

• fast-read-bold: a Chrome extension that bolds the start of each word, plus a word-by-word speed reader.
• MyShell: a Unix shell in C with pipes, redirection and background jobs.
• Wordle solver: picks each guess by Shannon entropy, with a Tkinter GUI.
• character.ai-style API: a team project in Spring Boot, MongoDB and Docker.

Intern, Electrolux Group · Bucharest · 2026

• Built GA4 funnel dashboards in D3.js with Python build scripts. GA4 names product categories differently when a product is viewed and when it is bought, so I joined the data on product model to make the funnel line up.
• Built a generator for Google Ads sitelink text that serves only approved copy and checks every benefit claim against the online store's terms and conditions.
• Built a crawler that collected over 13,000 pages and PDFs from Electrolux sites in 8 EU markets for a review under the EU's EmpCo greenwashing rules.

Intern, Nuvei

• Built a C# .NET microservice that takes XML requests over HTTP, validates them with an MD5 hash, logs every failure, and ships through CI/CD with unit tests.