muend.
AI engineering · LLM evaluation · agent systems Portfolio / 2026

Muhammed Enes Duran

I build AI systems that are measured before they are trusted.

Twelve systems along one line of work: LLM evaluation in production, deterministic benchmarks, a GPT-2 124M implementation, agent skills, guarded tool execution, an urban digital twin with a grounded assistant, simulation, reproducible data preparation and decision delivery. Open work links to something you can inspect; client work is described from its own records.

Focus
LLM evaluation, agent systems, MCP tooling, applied ML
Shipped
2 PyPI packages · Glama A · 2 Awesome lists
Method
Invariants first, failure modes second, artifact last
Based
Türkiye · open to collaboration
01

Systems

12 · inspect each system
923LLM eval cases built
18Agent skills
100Geoprocessing tools
124MGPT-2 implementation
00
Client work · LLM evaluation and data quality

EatBetter

Contract AI engineering on a nutrition app: a measurement platform, a cleaned food catalogue and a safer assistant, each change earned by evidence.

923benchmark cases
01
Deterministic evaluation · Exact Rust harness

benchfck

Generates its own machine-state tasks and scores them exactly, so no learned judge sits between a model and its result.

4instruction encodings
02
LLM foundations · GPT-2 implementation

neuron-to-gpt

From scalar autograd to a 124M-parameter GPT-2: byte-level BPE, checked pretrained imports and GPU training diagnostics.

124MGPT-2 parameters
03
Agent skills · GeoAI knowledge infrastructure

geoai-skills

Eighteen agent skills that make an AI assistant fail loudly on CRS, leakage, units and uncertainty instead of quietly guessing.

18specialist skills
04
MCP · Secure GIS execution

arcgis-mcp-bridge

Exposes ArcGIS Pro geoprocessing to AI agents while keeping the licensed ArcPy runtime behind a validated process boundary.

100geoprocessing tools
05
Applied simulation · LLM development game

Pixel LLM

A run-based game where each run trains one language model through ten stages, and every shortcut taken upstream shows up in the final model card.

10stages, data to launch
06
Urban digital twin · Studio platform

Mendirek / PlaceTwin

1,712 structures, 669 parcels and a team workspace: analysis remains linked to its source and confidence class.

1,712structures modelled
07
Applied simulation · Desktop game

FOUNDER.EXE

A startup simulation where cash, compliance, product and people interact under two real regulatory regimes.

26academy modules
08
Remote sensing · Reproducible data pipeline

sentinel-crop-pipeline

Takes Sentinel-2 scenes from search to training patches, counting every pixel it throws away along the way.

5auditable stages
09
Spatial DSS · Public-facing product

agri-dss

Turns regional agricultural data for 147 neighbourhoods into a crop plan a cooperative can pin to a village board.

147neighbourhoods
10
Data engineering · Web platform

Girişim Haritası

A source-linked company catalogue with repeatable collection, entity resolution, PostgreSQL and a quota-controlled API.

3,257company profiles
11
Desktop software · Document engineering

Plan Forge

A Windows report builder with local document parsing, versioned templates, expert approval and traceable DOCX/PDF output.

0.2.2Windows release build

Verified 2 November 2026. Source-linked project pages distinguish shipped products, research results and work in progress.

02

Experience

work · research · education
September 2026

AI Engineer (Contract) · EatBetter

Contract work on an AI-powered nutrition app, run on one rule: measure the model before changing it.

  • Evaluation platform. Built a 923-case benchmark from scratch: 569 human-adjudicated gold labels, a 277-case sealed holdout, 16 deterministic metrics and 742 automated tests in CI. Comparisons use user-clustered bootstrap confidence intervals, and the release gate counts missing evidence as a failure.
  • Model decisions by measurement. Retired a model that cost 7× more after showing no gain (ΔF1 −0.006, 95% CI [−0.023, +0.012]), and shut down a candidate pipeline that lost to the bare model in 4 of 4 comparisons.
  • Assistant safety. On a frozen 400-case scorecard, a fix for users with long memory profiles cut the missed-allergy-context veto rate from 96.4% to 2.1%, the only one of ten candidates to pass the gate. Five LLM-judge candidates were compared on 80 hand-labelled meals (best agreement 0.82 against a 0.66 rule baseline).
  • Production data quality. In a live PostgreSQL catalogue of about 387k rows, soft-deleted 88,970 and unpublished 90,735. Every write ran as a ROLLBACK rehearsal against a predicted row count before a human-approved COMMIT, and a rule plus LLM write gate closed the source.

Python · TypeScript / Node.js · PostgreSQL · Azure OpenAI

2026 – 2027

Undergraduate Researcher · TÜBİTAK 2209-A

A Sentinel-2 crop-classification study using U-Net semantic segmentation with spatially blocked validation. sentinel-crop-pipeline is the study's reproducible data backbone; its Urla run produced 990 training patches from 9 scenes.

Education

City & Regional Planning, with a second degree in MIS

B.Sc. City and Regional Planning, İzmir Institute of Technology (İYTE), 2021 – June 2027 (expected). B.Sc. Management Information Systems, Istanbul University (AUZEF), concurrent second degree, 2022 – June 2027 (expected).

Selected credentials: DataCamp AI Fundamentals certification (March 2026); Working with Hugging Face (October 2026), OpenAI API and prompt-engineering courses. Claude Academy / Anthropic: Claude 101 and AI Fluency course-completion badges (March 2026).

Open to collaboration

Build AI systems that hold up.

Available for collaboration around LLM evaluation, agent systems and MCP infrastructure, applied machine learning, data pipelines, decision-support products and applied simulations.

Based in İzmir, Türkiye · Collaboration scope and schedule agreed directly.

© 2026 Muhammed Enes Duran Set in Bricolage Grotesque, Instrument Sans & IBM Plex Mono Static site · no trackers
Technical diagram