Yaşar Enis Can

Project-based internship program (certificate of completion). Built Local RAG Assistant: a fully offline Q&A assistant on Microsoft Foundry Local — no cloud, no API keys, no data leaving the machine. github.com/eniscan17/local-rag-assistant

• Designed the full pipeline: chunking → embedding → SQLite vector retrieval → local chat model, with a Streamlit UI.
• Fixed a model-faithfulness failure (small models answering from general knowledge) with a deterministic relevance-threshold guardrail, verified by a 32-case automated eval harness.

The first steps of a bank's corporate credit analysis, automated on public filings of 10 BIST companies (112 reports). Numbers come from code, never from a language model: ratios from XBRL statements, with inflation-accounting restatement and trailing-12-month figures. A local LLM reads only the narrative notes (lawsuits, fines), and every claim it makes is checked against the cited PDF page.

• Ratios (net debt/EBITDA, interest coverage, liquidity, refinancing risk) validated against the companies' published figures; threshold breaches become early-warning flags.
• Net FX position read from the notes by layout-aware rules: 9/10 on a held-out year, where the LLM alone scored 7/10 by picking last year's column.
• Lawsuits and regulatory fines extracted by qwen2.5-7b and kept only if the quote and its figures appear on the cited page. On FY2024 reports labelled before the run, it found 4/4 material matters, including a 1.3 bn TL competition fine disclosed as a subsequent event.
• Generates a one-page Turkish credit summary per company, every figure linked to its KAP source page.

Independent follow-up to the internship project. On XQuAD — 1,190 Wikipedia questions professionally translated into both English and Turkish — the same questions are asked in both languages, so any gap comes from language alone. Every stage is measured separately, with 95% bootstrap confidence intervals.

• Turkish-aware hybrid retrieval (multilingual-e5 + stemmed BM25, rank fusion) raised Turkish top-1 retrieval from 74.5% to 89.5% with a 5x smaller embedding model.
• End-to-end answer evaluation (EM/F1) under four conditions — no context, gold passage only, new and old retriever — separated retrieval errors from generation errors.
• Switching to qwen2.5-7b raised Turkish answer F1 by 25 points (43.6 → 68.7); verified on the app's own test set (0% false refusals) before making it the default.

Two Foundry Local agents: a tool-calling agent that chains a calculator, Wikipedia, date/time and a local knowledge base, and a data-analysis agent that runs real pandas/matplotlib tools on uploaded CSVs instead of guessing numbers.

• Found that small on-device models labelled tool-capable often skip tool calls silently; fixed it with deterministic pre-routing and context injection.
• Shipped with Streamlit UIs, pytest suites and CI; findings and edge cases documented on GitHub.

A from-scratch PyTorch time-series pipeline predicting next-day stock returns — then, instead of stopping at "the model beats a baseline," I tested whether it would actually make money as a real trading strategy. It doesn't, and this section shows that on purpose: catching an overclaim before it ships is the point.

• Fixed a modeling bug across 3 iterations: predicting raw price lost to the naive baseline by 6-7x; predicting returns instead closed the gap.
• Backtested it as real trades with transaction costs: both models consistently lost to buy-and-hold — a validated negative result that changed the project's real conclusion, before any real capital was involved.