Software Engineer — Backend Systems, Retrieval & Applied AI
Building data-intensive backend services, hybrid retrieval pipelines, and deterministic evaluation tooling with ASP.NET Core, FastAPI, Rust, and PostgreSQL/pgvector.
SECTION 01
Language-agnostic dependency fingerprint and behavioral regression gate for AI agents in Rust.
ARCHITECTURE NOTE
Normalized model configs, prompts, MCP server capabilities, and tool schemas into RFC 8785 canonical JSON before SHA-256 hashing without importing agent code or relying on an LLM judge; eliminated false diffs from key reordering while catching silent behavioral drift.
Evidence-backed GitHub ecosystem analysis platform using hybrid retrieval and deterministic scoring.
ARCHITECTURE NOTE
Fused PostgreSQL full-text search with pgvector cosine similarity via Reciprocal Rank Fusion (RRF) and restricted LLMs strictly to query expansion while computing scores through versioned equations, backed by a durable PostgreSQL worker queue and 166 automated tests.
Local-first personal workflow platform unifying task planning, job aggregation, RSS, and full-text search.
ARCHITECTURE NOTE
Kept all user data strictly on local loopback while exposing bounded Streamable HTTP Model Context Protocol (MCP) tools alongside persistent PostgreSQL ingestion queues, an append-only event log, and an idempotent Chrome MV3 extension outbox.
Controlled evaluation of four RAG architectures across 422 legal documents and 1,030 questions (TUBITAK 2209-A).
ARCHITECTURE NOTE
Benchmarked four retrieval configurations over 5,979 chunks using paired statistical significance tests; reached 85.3% retrieval recall, 0.472 answer F1, and 0.790 composite RAGAS score with cross-encoder reranking, backed by a source-citing SSE Q&A prototype.
Offline-first IT support and ticketing prototype powered by a local LLM, pgvector retrieval, and rule routing.
ARCHITECTURE NOTE
Eliminated external LLM API calls to keep institutional ticket data strictly on-premise using a local LLM and PostgreSQL/pgvector; decoupled operational routing and triage policies into external Markdown rule files so support workflows update without redeploying application code.
Real-time YOLOv8 personal protective equipment detection prototype streaming from browser and phone cameras.
ARCHITECTURE NOTE
Decoupled live inference from fixed CCTV hardware during field validation by streaming frames directly from desktop browsers and mobile phone cameras into a YOLOv8 detection pipeline for access-control compliance checks.
SECTION 02
BOTAS (Petroleum Pipeline Corporation)
Developed iTicket, an offline-first IT support and ticketing prototype built with FastAPI, React, PostgreSQL/pgvector, and a local LLM without external API dependencies; implemented knowledge-base solution retrieval, request classification, and Markdown-driven operational routing rules maintained outside application code.
BOTAS (Petroleum Pipeline Corporation)
Developed a real-time YOLOv8 personal protective equipment (PPE) detection proof of concept for facility access-control scenarios, integrating live browser and mobile phone camera streams into a web/mobile prototype.
TUBITAK — 2209-A Research Grant
Benchmarked four RAG architectures across 422 Turkish legal documents (5,979 chunks) and 1,030 balanced evaluation questions for a TUBITAK 2209-A study; achieved 85.3% retrieval recall and 0.790 composite RAGAS score using cross-encoder reranking, and built a source-citing Q&A prototype with FastAPI, React, FAISS, BGE-M3, and SSE streaming.
SECTION 03
SECTION 04
Why dense embeddings miss exact technical tokens, and how to fuse PostgreSQL full-text search with pgvector cosine similarity using RRF in SQL.
Tracing high buffer reads in EXPLAIN ANALYZE and replacing a misordered index with a partial composite index to cut latency from 480ms to 11ms.