Göktuğ
Karaca

Reads the two browser tabs you last looked at, chunks and embeds them locally with Foundry Local, answers only from those chunks with visible sources, then drafts a new project from where the two topics intersect. Microsoft AI Innovators capstone.

A SaaS billing core built to stay correct under pressure: one idempotency key survives twelve concurrent retries, the request that lands exactly on the quota passes and the next one gets a 429, money is integer maths, and duplicate Stripe webhooks change nothing.

A five-step automation workflow, an MCP server that exposes it, and an agent built on top of both — written up as a case study that reports the honest result: the agent came out worse than the plain workflow.

A crawler that asks first: robots.txt before anything else, a rate limit it actually respects, schema-checked output, and logs you can read when a page changes shape.

Registration, login, token refresh and JWT-protected routes — the boring part of every backend, done properly once so it can be reused.

A movie and series tracker: accounts, watchlists, ratings and a browsing flow that does not fight the user. Internet Programming term project, deployed and public.

A car rental platform with dynamic pricing and a match advisor that narrows the fleet down to the two cars you probably want. Plain PHP and SQLite, deliberately no framework.

A containerized student information system with a C++ backend and PostgreSQL persistence — process handling, sockets and build discipline learned the hard way.

An intrusion detection pipeline over the CIC-IDS2017 dataset: cleaning, feature selection, model comparison, and a hard look at what the accuracy number is actually hiding.

Same 15×15 parking lot, two agents: one that plans with A* and one that learns with Q-Learning. The interesting part is where the learner wins and where it never catches up.

An 8-bit ALU, a 16-bit instruction decoder, an 8×8 register file and the wiring that makes them a CPU — plus the testbench that proves the ALU does what the spec says.

A library system used as an excuse to get object modelling right: a real role hierarchy, an automated fine policy, layered search, and a catalog kept thread-safe with a read/write lock.

A streaming platform modelled properly: 11 tables in 3NF, real constraints, an ER diagram that matches the schema, and fifteen queries from trivial to genuinely annoying.

A fridge that keeps track of what is inside and how long it has been there, so food gets eaten instead of thrown out. Embedded sensing on one end, computer vision on the other.

A personal finance app for Android: log what you spend, see where it went, and get a number that means something at the end of the month.

A very small CLI that picks your coffee brew from your mood and the time on the clock. Built in an evening, because not everything has to be a term project.