Projects — Denis Baldakov

Dissertation — Self-Hosted LLMs for Unit-Test Generation

My undergraduate dissertation: a multi-agent pipeline that generates unit tests for code, built to test whether small, self-hosted language models can rival a frontier model. A Programmer agent proposes a reference solution, a safety filter strips unsafe code, and a Tester agent writes the test suite — the same pipeline drives four quantised SLMs (Mistral-Small, DeepSeek-Coder, Llama-3.1, Nemotron-Mini) via LM Studio and Claude via the Anthropic SDK.

Benchmarked on HumanEval & MBPP with line coverage, mutation score and functional accuracy
Honest-by-default scoring: an empty-suite filter stops assertion-free tests scoring a trivial pass
Every run persisted as a JSON artefact; headline numbers come from a rescored 25-task evaluation

AI Mental Health Assistant

An end-to-end machine-learning pipeline for mental-health diagnosis and treatment support: data cleaning and EDA, model training and evaluation (including a clinicalBERT-based text model), delivered with a continuous-integration workflow, a written report, and the demo above.

Data preparation, exploratory analysis and train/val/test splits over a clinical dataset
Diagnosis and medication/therapy-prediction models, including clinicalBERT for clinical text
Automated test pipeline via GitHub Actions

More Projects

Open Source · Free app · v0.1.1

Presence — local-first job-search companion

A free, open-source desktop app that hunts for roles while you live your life. It reads published job-board APIs on a schedule, keeps a tracker with an honest timeline, sends a daily brief to your own Telegram bot, and prepares outreach through an interview-style agent session that shows where every sentence came from. Runs on your machine with a local model or your own API key under a hard daily budget — and never applies on your behalf: a human clicks every Submit.

Python 3.12FlaskSQLiteOllamaLLM agentsmacOS app

Code Download

Open Source · 7.2k★

HackerRank hiring-agent — Contributor

Diagnosed and fixed a PDF-extraction bug in HackerRank's open-source resume-screening tool that silently dropped whole sections of designed / multi-column resumes before parsing — contributed the fix with a regression test as an external pull request.

PythonPyMuPDFOpen SourceTesting

Pull Request Issue

NLP · Multi-Hop RAG

Prompt Engineering & Retrieval Strategies

Retrieval-augmented QA for multi-hop questions on HotpotQA: hybrid dense (Mistral + FAISS) and BM25 retrieval with cross-encoder reranking, comparing few-shot, chain-of-thought and self-consistency prompting. A group NLP project.

PythonMistralFAISSBM25RAG

Code

Deep Learning · Research

Deep Active Learning on CIFAR-10

Label-efficient image classification with a ResNet backbone, comparing selection strategies (TypiClust on supervised embeddings) and reporting accuracy vs. labeling budget.

PyTorchtorchvisionResNetActive Learning

Code Notebook

Computer Vision · Robotics

Object Detection Robot

ROS + LiDAR service robot that uses YOLO object detection to locate, collect and deliver items autonomously through a state-machine of navigation nodes.

PythonROSYOLOLiDAR

Code

Data Science

ML Feature Engineering

Feature engineering, exploratory data analysis and model comparison on tabular data, with reproducible evaluation notebooks.

Pythonpandasscikit-learnJupyter

Code Notebook

MLOps · Cloud

YOLOv5 + MLflow Microservices

Containerised MLOps platform: a YOLOv5 object-detection service with MLflow experiment tracking and MongoDB, orchestrated with Docker Compose and deployed to Kubernetes.

FastAPIYOLOv5MLflowDockerKubernetes

Code