AI / Backend Engineer — LLM applications, agentic systems, Python backends

Software engineering student (B.Sc., expected 2027) building LLM-powered backends end to end — architecture, API layer, model integration and deployment. Based in Istanbul.

LinkedIn · bilaleroglu3663@gmail.com

A conversational layer over an ERP API, built for Turkish SMB accounting workflows. ~10,000 lines of Python.

• LLM intent router on Claude Haiku classifying requests into 16 accounting intents — invoicing, e-invoice issuance, collections, waybills, counterparty records, reporting — dispatching each to its backend workflow
• Multi-tenant layer: per-tenant ERP credentials encrypted with Fernet, tenant-resolution middleware, 5-minute TTL cache to cut database round-trips
• Role-based access control across 4 roles, enforced in middleware at the request-path level
• OCR receipt ingestion (Tesseract + OpenCV), bulk Excel/PDF import with automatic counterparty resolution
• Per-transaction usage metering, JWT auth with bcrypt hashing, automated schema migrations, structured logging

Python FastAPI PostgreSQL SQLAlchemy Claude API Tesseract OCR Docker

A tool-calling agent that answers customer questions strictly from official product manuals and warranty documents.

• Two-stage retrieval: metadata-filtered search scoped to the user's registered product family alongside a general-document pass — preventing answers from one product model bleeding into another
• PDF ingestion pipeline tagging every page with product-family metadata before chunking and embedding into a FAISS index
• Product registration tool writing to PostgreSQL and automatically deriving warranty and maintenance dates
• Containerised with Docker, rebuilding the vector index at container start; Chainlit chat mounted on a FastAPI application

Python LangGraph LangChain OpenAI API FAISS FastAPI Docker

Autonomous mission software for a UAV that reached the finals of the TEKNOFEST International UAV Competition.

• 9-state finite state machine driving a complete mission: takeoff → lawnmower area scan → vision-based alignment → descent → payload release → target switch → return to launch
• ROS 2 vision node using HSV colour segmentation and contour corner analysis (approxPolyDP) to distinguish a red triangle from a blue hexagon in real time
• Sequential target masking so a completed target is ignored in later mission phases
• Proportional controller converging alignment to under 10 px of pixel error, velocity clamped at 0.5 m/s with a 5-second target-loss timeout
• AUTO ↔ GUIDED mode switching, programmatic waypoint generation and upload, NED-frame velocity commands over MAVLink

Python ROS 2 Humble ArduPilot SITL MAVLink Gazebo OpenCV

2D space shooter built in Unity with C#.

AI / LLM — Claude API · OpenAI API · LangGraph · LangChain · RAG · FAISS · pgvector · tool / function calling · intent classification

Backend — Python · FastAPI · REST APIs · PostgreSQL · SQLAlchemy · Supabase · Pydantic · JWT / OAuth2 · RBAC · multi-tenant architecture

Vision & Robotics — OpenCV · YOLO · ROS 2 · ArduPilot · MAVLink · Tesseract OCR

Infrastructure — Docker · Git · Linux