Tümay Turhan
I start with the problem then build toward a useful solution.
Researcher, builder, systems thinker: I build data and AI systems that make complex information easier to use.
SQL, Python, BI dashboards, machine learning and LLM/RAG systems — shaped through hands-on projects across analytics and AI.
About
I enjoy working with data. In every role I’ve had, I found myself drawn to the analytical side — making sense of numbers, building reports and figuring out what’s actually going on beneath the surface.
Over time, I realized this is where I do my best work: understanding data, finding patterns and turning complexity into something clear. I like clear thinking, honest signals and making things easier to understand.
Outside of work, I’m curious by default. History, philosophy, anthropology, video games — I’m drawn to how systems work and what drives people. I also like sharing what I learn. Writing, explaining and making complex things simpler is part of the process for me.
Structured, analytical and honest—that's how I think and how I work.
Domains
What I focus on
Understanding problems, processes, and what actually needs to improve.
Building simple models to explore patterns, predictions, and practical use cases.
Working with LLMs, RAG and AI workflows to make information easier to use.
Designing small, useful tools that people can actually understand and use.
Portfolio
Selected work
RiskHane - Istanbul earthquake risk pre-screening
Decision SupportProblem
Renting, buying or simply living in Istanbul often means making high-stakes housing decisions from fragmented earthquake-risk signals.
Why it mattered
The first decision is not a structural certificate; it is knowing when a neighborhood and building profile deserves expert attention.
Approach
Next.js frontend and FastAPI scoring API combining IBB scenario damage, neighborhood building stock, DASK region and user-entered building features into a relative risk band, score and explanation.
Key insight
The useful product is a responsible triage layer: clear enough for households, careful enough to show source boundaries and uncertainty.
Legal assistant — condominium & neighborhood law
RAG systemProblem
Legal knowledge exists—but everyday apartment and neighborhood disputes rarely get clear, grounded access; people fragment-search or defer to costly advice.
Why it mattered
Cost, conflict, and uncertainty compound. In law, hallucination cost is trust and liability—retrieval and grounding are non-negotiable.
Approach
Agentic routing to six collections; ChromaDB with legally structured chunks; FastAPI on Cloud Run, Streamlit UI, RAGAS + MLflow in the loop.
Key insight
Value is accessibility, not clever answers—grounded retrieval makes law usable, not only consultable.
Turkey Housing Market Dashboard
ELT + AI AgentProblem
Turkey's housing signals — price indices, mortgage rates, rent, FX — are scattered across sources with no unified macro view. In a high-inflation environment, price alone misleads.
Why it mattered
Affordability only becomes readable when multiple indicators move together. The gap between real and nominal spreads is where the actual signal lives.
Approach
Automated ELT pipeline pulling 20 macroeconomic indicators from TCMB's EVDS API into a three-layer Snowflake schema (RAW → STAGING → MART), refreshed monthly via GitHub Actions. A Claude-powered insight agent computes real vs. nominal price spreads and generates buyer / seller / investor signals. Frontend: Chart.js, FastAPI, deployed on Vercel.
Key insight
Automating data ingestion removes the bottleneck — the agent's job becomes interpretation, not collection. Structured signals replace narrative guesswork.
Steam marketplace analytics & forecasting
Business AnalyticsProblem
The market is large, but signal is weak—price and performance drivers hide behind noise and uneven data quality.
Why it mattered
Without an honest data layer, forecasts mislead. Weak signals must surface, not disappear, if decisions are to be trusted.
Approach
BigQuery + dbt, Prophet for growth, gradient boosting / RF for price drivers, Live dashboard for exploration.
Key insight
Not all signals deserve attention—Metacritic vs price is weak (r ≈ 0.23). Removing noise is as valuable as modeling signal.
YouTube comment intelligence
AnalyticsProblem
Comment threads hold product and sentiment signal; raw text does not scale to product or strategy decisions.
Why it mattered
Without a pipeline and a stable ontology, every analysis reinvents categories—BI and stakeholders never get comparable views over time.
Approach
Ingest via YouTube API into layered BigQuery tables, LLM enrichment for structured attributes, Streamlit for exploration and Live dashboard for reporting.
Key insight
Unstructured comments become useful only when mapped into consistent sentiment and topic signals—repeatable, decision-grade views for creators and sponsors.
GameKit — fast little tools for game night, no app download, just open and play
ProductProblem
Tabletop sessions constantly need dice, coin flips, and similar tools — but downloading a separate app, navigating ads, or carrying physical pieces all break the flow.
Why it mattered
A utility tool should open instantly and disappear into the moment. Friction before the first roll is friction that costs the game.
Approach
Web-based hub — no install, no ads, no account. Four tools in one place: coin flip, classic d6, RPG dice set, and Fate Chamber (AI-powered narrative decision tool on Gemini). React + Vite, ambient audio, deployed on Vercel.
Key insight
Speed and simplicity are the feature — a tool that loads in a second and does exactly one thing earns more trust than a polished app that asks for permissions first.
ThinkSlow — calm focus workspace
ProductProblem
Most focus tools demand attention—features, tracking, and pressure fragment attention further.
Why it mattered
Focus follows environment and cognitive load; without a calm, predictable space, attention destabilizes.
Approach
Ambient-first workspace: minimal UI, controlled sensory input, familiar scenes (rain, night, forest). Next.js, TypeScript, CSS Modules; bilingual UX; Vercel. State stays local—your sessions stay yours.
Key insight
Focus is regulation, not timers—environment shapes attention.
Brazilian e-commerce analytics
Data AnalyticsProblem
Public e-commerce data is easy to quote and hard to trust—analysts need clear numbers; leaders need profit context, not adjectives.
Why it mattered
One story, two surfaces: SQL work for rigor, then a small management dashboard anyone can read.
Approach
SQL project: clean joins, headline KPIs (~15.8M BRL revenue, high on-time delivery, ~6% repeat buyers). Dash: satisfaction translated into money on the page.
Key insight
Operations look healthy; loyalty does not—the dashboard's job is to show that gap in plain money terms.