Projects

Selected work

Six projects across retrieval systems, AI agents, multimodal systems, computer vision, forecasting and applied machine learning. Each is documented below as a short case study.

RAG Systems

QMS Assistant

A fully local, bilingual assistant over quality-management documents: cited answers, rule-based compliance checks, and a project lifecycle extracted from the documents themselves.

RAGQdrantBGE-M3Ollama

Open-source systemRead case

AI Agents 1st

Pulse

AI-orchestrated penetration-testing platform that runs twelve security tools and reports vulnerabilities on local models by default.

LangGraphOllamaNext.js

BirmingHack 2.0 · 1stRead case

Multimodal AI 1st

SocratEase

Multimodal speech-coaching platform scoring fluency, eye contact, and coherence from video via late fusion across three channels.

Multimodal AIXGBoostRoBERTa

BirmingHack 1.0 · 1stRead case

Computer Vision

Camera Calibration

A toolkit for calibrating cameras and removing lens distortion with OpenCV.

OpenCVJupyter

Research toolkitRead case

Data Science

Time-Series Forecasting

A full Box–Jenkins ARIMA workflow where every modelling decision is justified by a formal test rather than a heuristic.

ARIMAStatsmodelsBox–Jenkins

Open-source workflowRead case

Machine Learning

Config Performance Prediction

Four regression models compared across nine configurable software systems, with significance testing to back the ranking.

scikit-learnWilcoxonSciPy

Empirical studyRead case

Case 01

QMS Assistant

Retrieval-Augmented Generation · Local AI

Problem

Quality-management knowledge is spread across CMMI procedures, specifications and test plans, and the answer to a single question is often split between several of them. Sending those documents to a cloud model was not an option: to stay aligned with KVKK, Turkish data-protection law, everything had to run locally.

Approach

Hybrid retrieval in Qdrant merges dense and sparse BGE-M3 vectors with Reciprocal Rank Fusion, so exact terms and meaning both count. Documents are split into parent and child chunks, and late chunking pools each child's vector from a single pass over its parent, so small chunks keep their surrounding context. A local Ollama model answers at temperature 0, cites its sources, and falls back to an explicit "not in the documents" notice when retrieval is weak. Because BGE-M3 is multilingual, an English question retrieves from the Turkish corpus and is answered in English. Alongside the assistant sit a compliance checker, which runs deterministic structure and traceability rules before an LLM-as-judge content tier, and a lifecycle map extracted from the documents, where any claim that cannot be cited is dropped.

Challenges

The assistant occasionally answered from the wrong document. Logging the exact context the model received showed why: long documents whose relevance was spread across several chunks were ranking just below the top-three cutoff and never reached the model. A per-document scoring blend, the best chunk plus half the rest, let them surface without burying short, densely relevant ones.

Validation

A labelled compliance answer key deliberately includes compliant and reworded documents, so false positives are measured rather than ignored; the deterministic tiers score 1.00 precision and recall across 81 labels. Retrieval is tested with a 20-question bilingual set grouped by capability, including out-of-scope questions that must trigger the fallback. An early automated model comparison was discarded after review showed the LLM grader had never been shown the documents it was judging groundedness against.

Outcome

A one-command local system, with FastAPI serving both the API and a no-build React interface, backed by a pytest suite and running on a 16 GB laptop with no document or question leaving the machine. The repository ships with a synthetic corpus and an ingestion script for pointing it at real documents.

Case 02

Pulse

AI Agents · Cybersecurity

Problem

Penetration testing is largely manual, and most automated tools send target data to third-party cloud services — a non-starter in security-critical environments.

Approach

A LangGraph state machine fingerprints a web or repository target, then a planner node selects which of twelve industry-standard tools to run — nmap, sqlmap, dalfox, semgrep, bandit, trufflehog and pip-audit among them. An LLM reads the raw output, reasons across findings to build a MITRE ATT&CK–aligned exploit graph, and writes a structured vulnerability report. Reasoning streams live to a Next.js and FastAPI interface over Server-Sent Events. Ollama runs the models locally by default, with OpenAI and Claude available as optional backends.

Challenges

Coordinating non-deterministic model output with deterministic security tooling required strict state management and guardrails, delivered under a zero-data-leakage constraint and a 24-hour deadline.

Outcome

A local-first pipeline that runs end to end without sending target data off the machine, and won first place at BirmingHack 2.0 for Best Use of AI Agents on Arm.

Case 03

SocratEase

Multimodal AI · Speech

Problem

Public-speaking feedback is subjective and rarely available immediately after a session.

Approach

SocratEase analyses a speech video across three channels processed independently, then combined by late fusion. Vision tracks eye contact and facial expression using YuNet face detection and MediaPipe landmarks. Audio is reduced to fluency features — words per minute, lexical density, silence detection and MFCC coefficients via Librosa — and scored by an XGBoost model reaching a 93% F1. Text analysis covers coherence with RoBERTa-large, plus filler words, tonality and readability.

Challenges

Fusing audio, video and language models into one coherent score within a 24-hour build required careful scoping and a clean separation between capture, analysis and presentation. Late fusion kept each modality independently testable rather than entangling them in a single model.

Outcome

Built by a team of three, it won first place at BirmingHack 1.0 for Best AI Hack.

Case 04

Camera Calibration

Computer Vision · OpenCV

Problem

Every real lens deviates from the ideal pinhole model, introducing radial and tangential distortion that corrupts measurement and 3D reconstruction.

Approach

The toolkit detects the corners of a 9×6 checkerboard across many photographs and solves for the camera matrix and distortion coefficients, supporting both the pinhole and fisheye models. The recovered parameters undistort still images and video frame-by-frame.

Outcome

Reproducible Jupyter notebooks that calibrate a camera and undistort new footage, built as a prerequisite for computer-vision work at METU's Applied Intelligence Lab.

Case 05

Time-Series Forecasting

Data Science · Internship

Problem

Planning needs forward-looking estimates, not just historical reporting. But a forecast is only as trustworthy as the reasoning behind it, and ARIMA parameters are too often picked by habit rather than evidence.

Approach

An end-to-end Box–Jenkins workflow over an annual series of 63 observations spanning 1961 to 2023, where each decision is defended by a formal test. Stationarity is settled with Augmented Dickey–Fuller and KPSS in agreement — the level series is non-stationary while the first difference is not, establishing d=1. Model selection is a grid search over 36 ARIMA(p,d,q) combinations ranked by AIC, and rolling windows from 10 to 63 years are compared to weigh recency against sample size.

Validation

Residuals are checked against the white-noise assumption through time plots, a histogram against a normal overlay, Q–Q plots, autocorrelation analysis and a Ljung–Box test. The resulting one-step-ahead forecast carries a 95% confidence interval and is benchmarked against naïve baselines, so the model has to earn its complexity.

Outcome

A reproducible notebook that needs no external data download, with figures that regenerate on re-execution. The same forecasting technique underpins my ARIMA work on internal operational data at Doğuş Teknoloji.

Case 06

Config Performance Prediction

Machine Learning · Empirical Study

Problem

Highly configurable software exposes so many options that measuring the performance of every configuration is infeasible, so the performance of an unseen configuration has to be predicted. Comparisons between models are routinely reported on a single split, where an apparent winner may be nothing more than noise.

Approach

Four regression models — linear regression as a baseline, LassoCV, a multi-layer perceptron and a random forest — are trained on nine subject systems under a 70/30 split and scored on MAPE, MAE and RMSE.

Validation

Rather than declaring a winner from raw averages, differences are tested with Wilcoxon signed-rank tests at α=0.05 and corrected for multiple comparisons with Holm–Bonferroni, so only the differences that survive the correction are treated as real.

Outcome

A reproducible comparison across nine systems where every claimed difference is backed by a significance test, documented with a manual and a replication guide.