Technology
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.
The challenge
DSGENAI is an AI-driven platform that helps engineers generate and work through requirements for safety-critical systems. By the time I joined it, the platform had outgrown its original Flask architecture: it was hard to maintain, awkward to use, and its question-generation logic was tied to a single use case. More seriously, it carried critical data-leakage issues that are unacceptable in a safety context.
The platform needed to become something the research team could rely on and extend - stable, maintainable, generalisable to new use cases, and running on a current model.
The approach
I stabilised and optimised the DSGENAI platform end to end. I migrated the architecture from Flask to Streamlit, which made the codebase far more maintainable and gave engineers a cleaner, more direct user experience for the requirements workflow.
I generalised the question-generation logic so a single, well-structured engine could serve multiple use cases instead of being hard-wired to one, and I tracked down and fixed the critical data-leakage issues so the platform behaved correctly and safely.
Finally, I integrated GPT-5.1 as the platform's model, improving the quality of the generated requirements output and bringing the tool up to a current, dependable standard for the team's research.
Want similar results for your project?
Every project above started with a conversation. Let's figure out what yours needs.
Keep exploring
More projects.
Rail catenary pole placement automation
Weeks of expert engineering work, reduced to seconds.
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.
Integrated LoRa Sensor Monitoring & Analytics System
Turning raw LoRa telemetry from tree-mounted sensors into live dashboards that answer watering and growth questions - built in one week.
Revenue Administration MCP Server
An assistant that reads Turkish tax legislation from its official source at the moment you ask and answers with the article behind it - a system a certified acc
Reliability of LLMs in Safety-Critical Requirements Engineering
A controlled experiment measuring what an ungrounded, off-the-shelf chatbot contributes to safety-critical engineering.
Retrieval-Augmented Generation (RAG) Documentation Assistant
A RAG assistant that turns stacks of PDFs into a searchable knowledge base where every answer traces back to the source document.
LLM Prompt Optimization for Legal-Clause Classification
Comparative research that lifted F1 from 0.62 to 0.76 on Terms-of-Service clause classification through automatic prompt optimisation alone - without retraining
Smart Contract Analysis with NLP
An NLP system learns from Siemens' legal team's past contract revisions, flags the same clauses in a new contract, and proposes the edit that was made before.
Football Player Potential Prediction Model
A classification model predicting whether a player will be marked "highlighted" from 39 scout attribute scores - ROC-AUC 0.86 under 10-fold cross-validation.
Football Player Ranking System
A scoring engine that ranks players not by total score, but by how many attributes they exceed the statistical average for their own position.
Football Player Position Recommender System
A recommender that compares a player's attribute profile against the profiles of other positions and finds players could be more valuable in a different role.
Automated Connecting-Flight Optimization
A tool that recalculates every connection possible through the hub - day by day, with passenger volumes - when a single flight's time is shifted.
Flight Passenger-Count Prediction
A demand model predicting booked passengers on one-stop routes from a flight's own characteristics, selected by comparing eleven regression models.
Turkish Image-Captioning Benchmark on MS COCO 2014
A human-verified Turkish caption dataset covering all of MS COCO, plus five models trained on it - a new reference point for Turkish image captioning.
ESG Diversity & Sentiment Solution - CFA Poland Hackathon
An ESG prototype scoring gender diversity and news sentiment - 2nd place among 44 teams from 28 universities.
Automated ESG Scoring System - HackBogazici
An automated ESG scoring engine built from four scraped data sources - 2nd place among 14 hackathon teams.