Built, as part of a four-person team, an offline multilingual search tool for a Doctors Without Borders document library.
It is a single 33 MB SQLite file that needs only Python, and it is served over local Wi-Fi so it works without internet or a GPU.
Prepared the library at headquarters with Tesseract OCR for scans and photos and a local open-weight model (Mistral 7B via Ollama).
The model wrote summaries and multilingual keywords, added hypothetical questions and answers to the indexed passages,
and classified each document into an importance tier that is used as a ranking prior.
Retrieved with full-text search and an editable glossary, and added a local chat that answers with citations.
Recall@5 reached 0.9 in our evaluation.
The project was awarded Best Strategy at
Data 4 Good Challenge!
Retrieval and Grounded Generation over Scientific Papers
Built and evaluated retrieval over a corpus of ACL papers: a from-scratch TF-IDF baseline (MAP@10 0.373),
dense sentence embeddings (0.449), and two improvements on the dense baseline — structure-aware chunking
with top-3 weighted aggregation (0.489) and HyDE query expansion (0.548). I also tried rewriting queries with a LM and merging the results with reciprocal rank fusion, but this did not improve results.
Restricted grounding for the generator, to cite only permitted paper IDs, abstain when unsupported, and
validated it against abstention cases and prompt-injection attacks.
Multi-Agent Learning in Knights-Archers-Zombies
Developed a pipeline for a multi-agent environment that used a CenterNet detector to
extract zombie detections from visuals and incorporated handcrafted features into a behavior cloning policy.
Outperformed random (7%) and diagonal-shooting (1%) baselines, elevating the target hit rate to 16%.
Evolutionary Algorithm for Asymmetric Traveling Salesman Problem
Implemented an evolutionary algorithm that combines crossover/mutation operators with
optimization refinement, including stagnation handling and guided local search.
Outperformed a greedy algorithm by 21.3% on a 1000 cities scenario.
Analysis of Turkey's Grand Assembly
History
Investigated long-term trends in female parliamentary representation in Turkey by constructing
a multi-level dataset (national, regional, and city-level breakdowns) with 3,000+ records
across 21 legislative terms, scraped from official parliamentary archives using BeautifulSoup and Selenium.
Modeled and forecasted future female MP representation
using linear and polynomial regression with Scikit-learn,
selecting the polynomial model via R2 on held-out historical terms,
and testing forecast sensitivity to input window length (8 vs. 15 prior terms).
Airline Manager
Designed and developed a comprehensive full-stack
Node.js and React app for managing flight rosters,
including features for scheduling and ticketing.
Study Buddy App
Engineered a mobile app using Java SpringBoot and Android to facilitate the connection
between students who are studying similar courses, enabling them to collaborate and learn together.