Eyuphan Durmus — AI / ML Engineer. AI / ROBOTICSENGINEERING

OPEN TO ROLESTIRANA, ALRELOCATION OR REMOTE

Systems that see, measure and decide. Retrieval pipelines that show their evidence, inspection cells that measure to spec.

SELECTED SYSTEM / 01

JAGUAR

Robotic quality control, from laser scan to digital twin

~40s
scan to report, per part
17
features measured against tolerance
<10%
gage R&R — the band where a measurement system is accepted outright

An automated inspection cell at MND Robotics whose software stack I owned: multi-angle laser scans become a dense 3D point cloud, registered into a digital twin, and customer-defined features are checked against their tolerance limits.

Open3DCloudCompareSciPyOpenCVMech-EyeDockerUbuntuReactFlaskSentry
OPEN DEMO

SYSTEM ARCHITECTURE

SENSOR
Mech-Eye Laser
ROBOT
Cobot A* path
ANALYSIS
Open3D SciPy
OUTPUT
Dashboard QC link

IN PRODUCTION

A React and Flask console runs the cell: cycle control, live point-cloud preview and the QC overlay. Operators use it every day.

INSPECTION CELLMND ROBOTICS
The robot arm carrying the scanner head inside the Jaguar inspection cell

Customer parts digitally removed under NDA.

CELL SPECIFICATION
CYCLE
~40s
OUTPUT
3D digital twin
EQUIPMENT
Mech-Eye LSR QC link

SELECTED SYSTEM / 02

75 ENDPOINTS INDEXED · 6 STEPS · 4 WALKED

INVESTIGATION AGENT

A support case in, an evidence trail out, and a report that names what it could not prove

10 HARDEST OF 52 REAL CASES · MEASURED ON REPLAY · 2 SHOWN BELOW

STEP 6 / 6NO CALL — DECIDES WHAT IT CANNOT PROVE

Calls nothing. Every cause the panel can test is ruled out, and the one signal left is circumstantial — so it refuses to call the cause confirmed, names the gap, and escalates.

Each point is an endpoint; the lit path is the order this case walked them

CAUSES IT TESTED

6
  • The coupon simply expired — RULED OUT.
  • Its value was consumed by trading losses — RULED OUT.
  • Cancelled by a completed balance transfer — RULED OUT.
  • Cancelled by a rejected transfer attempt — PLAUSIBLE, UNTESTABLE HERE.
  • Manual cancellation by an operator, or an abuse disqualification — STRONGLY SUPPORTED.
  • The campaign behind it was terminated — UNRESOLVED.

3 RULED OUT · 1 STRONGLY SUPPORTED, NOT CONFIRMED · 2 LEFT OPEN

6
steps in the loop before it stopped
2 of 6
checklist items it could not prove
4
distinct endpoints it had to walk

An AI investigation agent for a crypto exchange's support and finance escalations. A raw case in Turkish or English goes in; it works the case across as many read-only endpoints as it takes, and closes a checklist item by item on the rows it actually fetched. It was measured by replaying the ten hardest of 52 escalations support and finance had already resolved by hand, the ones actually worth testing, so every run could be judged against an answer that was already known. One of them had taken the dev team two developer-days; the agent worked it in minutes, for about $0.03 in model spend.

ChromaBM25RRFnomic-embed-textOllamaOpenRouterFastAPIPython

SYSTEM ARCHITECTURE

PLAN
Next query from the gaps
EXECUTE
Routed search, guarded call
EVALUATE
Is the evidence enough?
REPORT
Names what it could not prove

PROVENANCE

Every merged fact is stamped with the endpoint it came from, so same-named fields from different endpoints stay separate facts, never silently collapsed into one.

WHAT IT REFUSES TO DO

G2NEEDS INPUT

A missing context field stops the call. It comes back with the field it needs instead of guessing a value and answering confidently from the wrong rows.

13CHECKLIST GATE

The loop cannot declare itself done while a checklist item is open. Whatever is still open is named as a gap rather than quietly dropped.

16CLAIM ADJUDICATION

A user's figure is judged only against the primary record for its own field. Derived accounting cannot refute it; at most it leaves the claim unresolved and says so.

REPLAY EVALUATION

The ten hardest of 52 escalations the team had already closed by hand, replayed so each run could be scored against a known answer.

5
reached the same conclusion support and finance had reached by hand
4
said the data was not there and escalated, instead of inventing a cause

On one it over-read: it named a cause the evidence did not support. That case is in the log below.

ACADEMIC PROJECT

EEG OCULAR ARTIFACT DENOISING

Deep learning for biomedical signal processing

B.Sc. thesis: a U-Net deep learning pipeline that removes EOG artifacts from 19-channel EEG signals, clean enough for downstream analysis.

U-NetTensorFlow/KerasSciPy
EEG Ocular Artifact Denoising — result 1 of 3
EEG Ocular Artifact Denoising — result 2 of 3
EEG Ocular Artifact Denoising — result 3 of 3

EXPERIENCE

FROM R&D
TO APPLIED AI

Robotics inspection cells, manufacturing systems and retrieval-heavy agent platforms.

  1. NOV 2025 - AUG 2026

    AI/ML ENGINEER

    — DARKEX (CRYPTO EXCHANGE), TIRANA, ALBANIA
    • Rebuilt the support and finance escalation workflow around a centralized Zendesk hub, the daily entry point for two teams of 30 people, cutting average resolution time by 40%.
    • Built and evaluated an AI investigation agent on top of it, with hybrid retrieval and multi-LLM routing across 75 endpoints and 29 categories, and measured it by replaying the ten hardest of 52 already-resolved escalations against their known outcomes. One that had taken the dev team two developer-days ran in minutes, for about $0.03 in model spend.
    Chroma, BM25, RRF, Ollama, OpenRouter, OpenAI, Zendesk, Python, FastAPI, Redis
  2. AUG 2024 - NOV 2025

    R&D ENGINEER

    — MND ROBOTICS, MANISA, TURKEY
    • Owned the entire software stack for Jaguar: multi-angle laser scanning, point-cloud reconstruction, A* scan planning and the measurement pipeline, producing a 3D digital twin in roughly 40 seconds per part. Hardware and mechanics by the MND team.
    • Built the React-Flask operator console for real-time control and monitoring, integrated with external QC equipment. In daily use on the shop floor.
    • Migrated a Turkish Aerospace Industries AGV from ROS 1 to ROS 2 and supported deployments across customer sites.
    ROS 2, Open3D, CloudCompare, OpenCV, Mech-Eye, Docker, React, Flask, Sentry
  3. AUG 2023 - SEP 2023

    R&D INTERN

    — IMECAR, ANTALYA, TURKEY
    • Prototyped ROS 2-based software for an autonomous waste-collection robot.
    ROS 2, Python
  4. JUL 2023 - AUG 2023

    SOFTWARE DEVELOPER

    — DT CLOUD, ANTALYA, TURKEY
    • Built and deployed a full-stack ML classification pipeline: React frontend, Flask backend, a PostgreSQL data layer, and JWT-based authentication and session handling.
    React, Flask, PostgreSQL, JWT
  5. EDUCATION
    GRADUATED 2023

    B.SC. ELECTRICAL & ELECTRONICS ENGINEERING

    — AKDENIZ UNIVERSITY, ANTALYA, TURKEY

    Thesis: U-Net ocular-artifact denoising for 19-channel EEG.

SKILLS

RETRIEVAL & AGENTS

  • Hybrid retrieval
  • BM25
  • RRF
  • ChromaDB
  • RAG
  • Tool calling
  • Ollama
  • OpenRouter
  • OpenAI
  • LLM evaluation

DEEP LEARNING

  • TensorFlow/Keras
  • U-Net

ROBOTICS & 3D

  • ROS 2
  • Open3D
  • OpenCV
  • Point-cloud processing
  • CloudCompare

APPLICATION & WEB

  • Python
  • FastAPI
  • Flask
  • React
  • TypeScript
  • PostgreSQL
  • Redis
  • Socket.IO

TELEPHONY & REAL-TIME

  • Asterisk (AMI)
  • Issabel
  • Event-driven asyncio

Contact-centre calling platform: sole engineer, zero to production on a live Issabel/Asterisk PBX.

PLATFORM & INFRA

  • Docker
  • Linux
CERTIFICATIONS
Machine Learning SpecializationDeepLearning.AI & Stanford Online (Coursera) - Dec 2024View credential
LANGUAGES
TURKISH- Native
ENGLISH- Professional working proficiency