name:      Yağız Akkaya
role:      Computer Engineering Student @ METU NCC
focus:     [ LLM Tool Calling & Agents, Retrieval-Augmented Generation, Computer Vision, Edge AI, Backends ]

I build systems that see and systems that read — computer vision models that run on constrained edge hardware, and LLM applications that let people ask questions of documents and databases they'd never have time to open. Lately that means tool-calling agents: the model picks a typed tool, the database or the document does the real work, and code checks the answer before anyone sees it.

Most of my work lives at the awkward, interesting seam between a model that works in a notebook and a system that survives real production environments: single-label datasets meeting multi-defect reality, detector jitter breaking naïve logic, GPU memory limits under concurrent ingestion, and high-throughput vector indexing. I measure what I ship with gold-set evaluations (held-out splits, honest failure counts) instead of quoting the best demo run.

🚁 "CENGIZ" — Unmanned Coaxial Helicopter Helicopter Design Competition · 75.3 pts	🚁 Modular Multipurpose Helicopter Helicopter Design Competition · 78 pts
Unmanned coaxial helicopter for disaster-area connectivity and marine monitoring. I led the rotor configuration analysis and vehicle pre-sizing, and built the vehicle's 3D model entirely through Blender Python scripting — modular components with a game-optimized export, shown as a playable GTA V mod (Yavuzkopter). Aerospace UAV Design Systems Engineering Blender Python	One airframe, three variants (search & rescue, commercial, VIP), designed with a 6-person team around a modular architecture. Aeronautics Modular Architecture

Sep 2026 · On-site

• Kervansaray — Hotel Parking License-Plate Recognition & Query System: Built an end-to-end data-management system for fixed-camera hotel parking entry/exit control, from synthetic data generation to natural-language querying.
  • Backend & Data Pipeline: Engineered a Flask + PostgreSQL 16 event store (Docker, SQLAlchemy, Alembic) with idempotent ingest, plate canonicalization, bounded fuzzy matching, entry/exit session reconciliation, and a deterministic synthetic data generator.
  • LLM Query Layer: Turkish natural-language queries answered through tool calling over typed SQL tools (NVIDIA NIM Gemma 4 31B, Gemini Flash fallback); the database does all counting, the model only picks the tool and phrases the result.
  • Evaluation: Built a 209-question gold set (56 dev + 153 held-out) scored against an independent Python oracle with Wilson 95% confidence intervals; a patch-ablation study showed the model no longer needed the old deterministic fixes (52/56 → 53/56 on the dev set without them).
  • Production Readiness: Deterministic rule engine with Server-Sent Events alerts, Prometheus observability, structured logging, and a live public demo at kervansaray.yagizakkaya.com.tr.
• Inventory Verification — Fixed-Camera Object Detection & State Control: Developed a workspace inventory-verification system for defense-industry use cases such as ammunition tracking and assembly-line control.
  • Detection & Logic: Fine-tuned a YOLO11 model paired with a finite state machine to verify scene state while filtering out detector jitter.
  • Data Labeling: Hand-labeled 50 of 452 frames, then bootstrapped a model to auto-label the rest, cutting labeling time significantly.
  • Edge Deployment: Deployed to NVIDIA Jetson with TensorRT INT8 quantization for real-time inference using a Luxonis OAK-D camera.
• Research & Web: Prototyped flawscope, a two-stage hybrid defect-detection pipeline (PatchCore/PaDiM anomaly detection + CNN and WinCLIP branches fused with RRF) for aircraft-engine part metallography across 12 defect types, and deployed a personal portfolio site (yagizakkaya.com.tr) with Caddy reverse-proxy routing and visitor analytics.
• Technologies: Python Flask PostgreSQL SQLAlchemy NVIDIA NIM Gemini YOLO11 OpenCV NVIDIA Jetson TensorRT Luxonis OAK-D PatchCore CLIP Docker Caddy

Apr 2026 – Jul 2026 · Remote

• Vector Ingestion & Search Sync Pipeline: Engineered a high-throughput legal document ingestion and indexing pipeline to synchronize millions of judicial decisions into Qdrant and Meilisearch databases.
• Database-Driven Ingestion: Engineered a concurrent search sync engine (--from-db mode) using PostgreSQL row-level locking (select_for_update) to support parallel worker execution without resource collision.
• VRAM & Speed Optimization: Developed a shared OOM-safe cascading encode algorithm (full → half → quarter → one-by-one) and single-document batching, reducing padding waste and increasing embedding speed by ~30%.
• Fault-Tolerant Outage Recovery: Built connection-outage resilience that automatically releases claimed documents back to pending_sync during target database failures and halts after 3 consecutive errors to prevent CPU spinning.
• Process Lifecycle & OS Management: Integrated process signal handlers (SIGINT/SIGTERM/SIGBREAK) and resolved a Windows-specific process zombie handle bug using ctypes for clean database lock releases on shutdown.
• Live Observability: Integrated Qdrant server operation_id tracking into Django WebSockets for real-time progress logging and dashboard pipeline statistics monitoring.
• Technologies: Python Django ORM PostgreSQL Qdrant Meilisearch SentenceTransformers CUDA PyTorch WebSockets

Jan 2026 – Feb 2026 · On-site

• CilekAI — Enterprise RAG Document Assistant: Designed and built a production-grade RAG system enabling 500+ factory employees to query internal documents (PDF, Excel, images) with inline citations.
• Multi-LLM Orchestration: Integrated 5 LLM providers (Groq Llama 3.3, Google Gemini, GPT-4o, local Ollama) behind an automatic fallback chain to ensure near-100% service uptime.
• Hybrid Retrieval & Reranking: Engineered a 4-stage search system combining FAISS semantic search and BM25 keyword matching with RRF score merging, Qwen3-Reranker cross-encoder, and Turkish morphological normalization, boosting accuracy from 0.15 to 0.95.
• Dual Vector-Store Architecture: Deployed a hybrid vector database using server-side Qdrant and in-process FAISS, utilizing incremental indexing based on file modification times (mtime) and SHA-256 file deduplication.
• Performance & Semantic Caching: Achieved <10ms FAQ response times and reduced LLM API costs by ~30% by implementing a FAISS-backed semantic caching layer and Levenshtein-based fuzzy matching.
• Enterprise Security & Architecture: Designed a 5-layer Clean Architecture with department-based role-based access control (RBAC), Prometheus-based observability metrics, and Docker Compose multi-service deployment.
• Technologies: Python Docker Compose Qdrant FAISS Ollama Qwen3 BM25 Levenshtein Prometheus Clean Architecture

Middle East Technical University, NCC — Bachelor of Science in Computer Engineering
2021 – Present

• Relevant Coursework: Algorithm Analysis, Database Management and File Structures, C Programming, Data Structures, Python Programming, Linux Basics, Logic Design, Discrete Computational Structures.
• Projects & Competitions:
  • Senior Design (ESC 491/492, CNG + EEE): UAV threat map and decision-support system with a physical demo; I own the RAG-based doctrine knowledge base, evaluation and the report — see uav-cuas-datasets.
  • TEKNOFEST 2025 Helicopter Design Competition: Unmanned coaxial helicopter "CENGIZ" (Graded 75.3, Ranked 5th place in Türkiye).
  • TEKNOFEST 2024 Helicopter Design Competition: Modular multipurpose helicopter (Graded 78, Ranked 4th place in Türkiye).

• 🛰️ Senior design: building the doctrine knowledge base (RAG with mandatory citations) for our UAV threat map, heading to a live single-camera field demo with a mock drone
• 🧪 Writing evaluations before features — gold sets, held-out splits and prompt-injection checks for every LLM app I ship
• 🤖 Moving from fixed RAG pipelines to tool-calling agents that read documents and databases through typed tools
• 🔬 Still pushing edge inference — TensorRT INT8, measuring the accuracy/FPS curve honestly instead of quoting the best number

📫 Let's talk

yagizakkaya10@gmail.com

"Anyone can train a model. The engineering is everything that happens after."