Computer Engineer
Emir
Kahraman
I build AI systems, distributed backends and security tooling — from RAG/LLM applications to a Rust security proxy, a Kafka-driven distributed game engine and a realtime telescope-control simulation.
01 About
Computer Engineering student specializing in AI systems, data analysis and full-stack development. Hands-on experience designing LLM/RAG applications and AI-agent tooling, writing analytical SQL over relational schemas, and building Python/FastAPI and React/TypeScript services.
- Education
- B.S. Computer Engineering, Ege University — CGPA 3.60, 2021–present
- Experience
- IT Intern, Goodyear — software prototyping & industrial IT, Aug 2025
02 Selected work
-
- Problem
- Turning piles of PDFs, scans and images into presentations is slow manual work — and raw LLM output can't be trusted to stick to the source material.
- Solution
- OCR ingestion feeds a local RAG pipeline over ChromaDB, so generated slides stay grounded in retrieved evidence; pluggable LLM providers (Vertex AI or local Ollama) draft the deck, exported as PowerPoint or PDF.
- Build
- Asynchronous FastAPI backend for ingestion, retrieval, slide planning and export; React + TypeScript front end with a separate slide-rendering layer for preview/export consistency.
B.S. thesis project — codebase private; architecture walkthrough and live demo available on request.
-
- Problem
- Terminal output an AI agent reads is untrusted input: it can leak API keys and credentials into the model, or hijack the agent through prompt injection.
- Solution
- A single Rust binary proxies every command through a nine-stage pipeline — execute, ANSI-strip, optimize, truncate, scrub, wrap, audit — so the model only ever sees sanitized output inside untrusted-data markers.
- Proof
- Pattern + entropy-based secret redaction, multilingual injection neutralization with a fail-closed
--strictmode, defanged markers, private-by-default audit logs — verified with property tests and benchmarks.
-
- Problem
- A two-player strategy game where neither player — nor a crashed server — may corrupt the game or peek at the opponent's hidden state.
- Solution
- Every order is validated through Kafka before touching state; the engine runs as three Go instances (leader + followers) on a three-broker KRaft cluster, with state in a log-compacted
game.sessiontopic rehydrated on restart. - Proof
- 13 Avro schemas across 10 topics behind Schema Registry, a 13-step turn pipeline, information hiding enforced at every layer — 205 Go tests, race-detector clean, one-command Docker setup.
-
- Problem
- How does raw source code become executed output? Star Language answers it end to end, from scratch, in C.
- Solution
- A lexical analyzer scans
.stasource into a token stream, then a tree-walking interpreter executes it —intandtexttypes, arithmetic, string operations, loops and console I/O. - Proof
- The entire toolchain reads end-to-end in a single C source file — a compact reference for how interpreters work.
-
- Problem
- Marketplace rules — stock, billing, customer tiers, returns — break silently when application code is the only thing enforcing them.
- Solution
- A multi-vendor relational schema covering users, shops, products, orders, payments, logistics, reviews and returns, with business rules enforced in-database through PL/pgSQL triggers.
- Proof
- Views and stored procedures ported across PostgreSQL and SQL Server; analytical SQL including RFM customer segmentation with window functions.
03 Experience
IT Intern
Goodyear Tire & Rubber — Adapazarı plant
Aug 2025
- Shipped four software prototypes end to end during a 20-day industrial IT internship, closing with a live technical demo.
- Prepared port and VLAN migration for a SAP HANA switch installation on production network hardware over the switch CLI.
- Completed Cisco Packet Tracer labs — routing, DHCP, DNS, VLAN, WPA2 wireless — plus STP, firewall and VPN fundamentals in a live server room.
Prototypes
-
- Python server computes live ephemeris for planets, moons and deep-space objects (Skyfield + Astropy, Hipparcos/Gaia catalogs) and converts RA/Dec to Alt-Az with refraction and parallax corrections.
- Targets stream to Unity over a custom UDP protocol — JSON packets with checksums, retransmission and timeouts — driving step motors simulated at 0.01125°/step with 10:1 gearing, 16× microstepping and shortest-path slewing.
- Closed the internship with a live demo tracking multiple celestial targets in real time.
-
- Room system with shareable codes, turn synchronization, instant letter/digit feedback and reconnect handling; letter-indexed Turkish dictionary for fast word validation.
- Bot opponent prunes the candidate space by consistency filtering and picks each guess by minimax group reduction.
- Profiled with Chrome DevTools — guess selection cut from O(n³) to O(n²), heavy scoring moved to Web Workers, memory tamed with chunking and object pooling.
-
- NumPy math engine evaluates coefficients and computes roots and the vertex; Matplotlib renders smooth high-resolution curves inside the Tkinter UI.
- Regex-based input validation with clear error messages — bad input warns instead of crashing.
- Tested across coefficient ranges and reviewed for readability and modularity.
04 Stack
Languages
Python · SQL · TypeScript · Go · Rust · C# · Java · C
AI & data
RAG · ChromaDB · Qdrant · OCR · PostgreSQL · T-SQL
Backend & realtime
FastAPI · Node.js · React · Kafka · WebSocket · UDP/TCP · Unity · Docker · Git
05 Contact