All projects

Other

Rail catenary pole placement, automated

Days of expert engineering work, reduced to seconds.

Days to seconds · Engineering time

A rail infrastructure engineering firm August 2026 Workflow Automation & Custom Software Featured
Automation Python LLM Custom Software
Rail catenary pole placement, automated

The challenge

A 148 km high-speed rail line needed 4,351 electrical (catenary) poles placed along its length. The spacing rules are anything but fixed: poles can sit up to about 65 m apart on straight track, but tighten to as little as 27 m on sharp curves, with the exact maximum set by curve radius and track cant read from a 36-page lookup table.

Nothing can be placed on a structure. On this line alone that meant avoiding 159 structures across 166 track segments: 126 culverts, 10 viaducts, 9 overpasses, 8 tunnels and 4 bridges.

Done by hand, the job means cross-referencing a spreadsheet, a 36-page PDF table and a structure list, then marking every position on a drawing, kilometre after kilometre for 148 km. It takes days, is easy to get slightly wrong, and starts over from scratch whenever the line design changes.

The approach

I built a Python system that does the entire job from the firm's own existing files, with no new data entry.

  1. It reads the track geometry (segment start and end, element type, curve radius and cant) straight from the engineering spreadsheet.
  2. It uses an LLM to parse the 36-page maximum-spacing PDF into structured, cached data, replacing the manual lookup that used to slow everything down.
  3. It derives the forbidden zones from the structure sheet, handling a tricky quirk where consecutive rows mark a structure's start and end with no explicit end column.
  4. It places every pole with a greedy algorithm that walks the line and pulls the spacing in tighter as curves and structures demand.
  5. It converts kilometre positions into real X/Y coordinates and outputs an Excel file for review plus DXF/DWG that open directly in AutoCAD.

Correctness was proven, not assumed. The output was compared directly against 4,351 pole coordinates the firm's own engineers had placed by hand as ground truth, backed by 68 automated tests covering geometry, structure handling, coordinate conversion, placement logic and output format.

Stack: Python, a greedy placement algorithm, LLM-assisted PDF parsing, and Excel/DXF/DWG output.

The outcome

Results that moved the needle.

Days to seconds

Engineering time

A multi-day manual task became a single command.

4,351

Poles validated

Checked against the firm's hand-placed reference points.

68

Automated tests

Covering geometry, placement logic and output format.

148 km

Line automated end to end

Want similar results for your project?

Every project above started with a conversation. Let's figure out what yours needs.

Keep exploring

More projects.

Healthcare · Decoto International

D-Risk - MedTech Marketplace with AI Company Profiling

A three-sided marketplace linking medtech startups with investors and specialist freelancers - matched through AI document profiling.

View
Technology · Ottronic

Integrated LoRa Sensor Monitoring & Analytics System

Real-time monitoring and analytics for LoRa environmental sensors - from raw ingestion to geospatial drought and growth insight.

View
Other · Independent project

Revenue Administration MCP Server

A retrieval-only MCP server answering Turkish tax and regulation questions strictly from official government sources - zero hallucinations.

View
Technology · Fraunhofer IKS, Munich

Reliability of LLMs in Safety-Critical Requirements Engineering

Master's-thesis research measuring how LLM assistance affects correctness, efficiency and trust in safety-critical requirements engineering.

View
Technology · Fraunhofer IKS, Munich

DSGENAI - AI Safety Requirements Engineering Platform

Stabilising and modernising an AI-driven safety-requirements platform - Flask to Streamlit, GPT-5.1, and critical data-leak fixes.

View
Technology · Independent project

Retrieval-Augmented Generation (RAG) Documentation Assistant

A Gemini-powered RAG assistant that turns PDF libraries into an accurate, source-grounded knowledge base.

View
Technology · Technical University of Munich

LLM Prompt Optimization for Legal-Clause Classification

Prompt-engineering research that lifted legal-clause classification accuracy by up to 20% on Terms-of-Service documents.

View
Technology · Siemens

Smart Contract Analysis with NLP

An NLP system that reviews business contracts and recommends changes to keep them compliant with company policy.

View
Other · Scoutium

Football Player Potential Prediction Model

A machine-learning model that reads scouting attributes and predicts player talent level with roughly 85% accuracy.

View
Other · Scoutium

Football Player Ranking System

An attribute-based scoring engine that ranks 750+ players and rewards standout attributes with a statistical bonus threshold.

View
Other · Scoutium

Football Player Position Recommender System

A recommender that analyses 750+ players' attributes by position and surfaces players who would excel in an alternative role.

View
Other · Turkish Airlines

Automated Connecting-Flight Optimization

An automated system that recomputes viable connections across 1000+ flights the moment schedule times change.

View
Other · Turkish Airlines

Flight Passenger-Count Prediction

A scikit-learn regression model that predicts passenger counts per flight from operational flight features.

View
Technology · Istanbul Bilgi University

Turkish Image-Captioning Benchmark on MS COCO 2014

A large-scale Turkish image-captioning benchmark - 616,767 human-verified captions and five trained vision-language models.

View
Finance & Fintech · CFA Poland (hackathon)

ESG Diversity & Sentiment Solution - CFA Poland Hackathon

An ESG prototype scoring gender diversity and news sentiment - 2nd place among 44 teams from 28 universities.

View
Other · Boğaziçi University IT Club (hackathon)

Automated ESG Scoring System - HackBogazici

An automated ESG scoring engine built from four scraped data sources - 2nd place among 14 hackathon teams.

View