Portrait of Denis Craciun

Denis Craciun

AI Engineer | Football Enthusiast | Muay Thai Practitioner

View CV Get in Touch

About Me

AI Engineer focused on shipping production-grade AI systems: agentic RAG, multi-agent orchestration, MCP-based tooling, and fine-tuned open-source LLMs. I work end-to-end across the stack — Python and PyTorch, LangChain/LangGraph and Hugging Face, vector databases (Pinecone/Chroma), QLoRA fine-tuning, ETL pipelines, model evaluation and hallucination mitigation — turning research into practical solutions, from tourist chatbots to self-improving agents. I’m driven by continual learning at the intersection of AI, software engineering, and finance.

AI Engineering Skills & Stack

Languages & Frameworks

  • Python
  • PyTorch
  • TypeScript
  • LangChain / LangGraph
  • Hugging Face Transformers
  • LlamaIndex

LLMs & Agents

  • Agentic RAG
  • Agent Orchestration
    • ReAct
    • Plan-and-Execute
    • Supervisor / Multi-agent
    • Reflection & Self-critique
    • Routing & Prompt Chaining
  • MCP (Model Context Protocol)
  • Tool / Function Calling
  • Agent Skills
  • Sandbox Execution
  • Persistent Agent Memory
  • Goal Decomposition & Task Graphs
  • Self-Improving Eval-Gated Loops
  • Prompt Engineering

Retrieval & Data

  • Vector DBs (Pinecone / Chroma)
  • Embeddings & Reranking
  • Hybrid Search
  • Chunking Strategies
  • Synthetic Dataset Generation
  • ETL Pipelines

Fine-Tuning & Optimization

  • QLoRA / LoRA Fine-tuning
  • PEFT · TRL · bitsandbytes
  • Distributed Training (DeepSpeed ZeRO)
  • Quantization & KV-Cache Compression
  • Adapter Merging
  • Training LLMs from Scratch
  • Hugging Face Hub

Inference & Deployment

  • Self-Hosted Serving (Ollama / vLLM / llama.cpp)
  • OpenAI-Compatible Endpoints
  • GGUF Conversion & Modelfiles
  • Chat Templating
  • Inference Benchmarking
  • Containerized Workloads (Docker)
  • FastAPI

Evaluation & Reliability

  • NLP
  • Model Evaluation (Evals)
  • Hallucination Mitigation
  • Guardrails
  • Observability & SLOs
  • Weights & Biases

Projects I Worked on

sapiens project

sapiens

A Claude Code skill that explains AI/ML and LLM-engineering code clearly — to the point, in plain language, grounded in your codebase rather than generic snippets. The evolutionary next step from "caveman": where caveman compresses to save tokens, sapiens clarifies so the code is actually understood.

Claude Code Agent Skills LLM
View Project
omni project

omni

A durable, observable, self-improving agentic operating system you talk to — with zero dependencies. Omni turns natural-language requests into verified work: it decomposes goals into tasks, runs specialized workflows, keeps durable memory, governs risky actions, and improves itself behind an evaluation suite. Built entirely on the Node standard library.

Node.js Agents SQLite
View Project
DerpGPT project

DerpGPT

DerpGPT is a conversational 50M parameters AI model that generates responses based on synthetic persona-chat datasets. It was trained on multi-turn dialogues but often struggles to keep things coherent. Whether it's making up bizarre responses, misunderstanding context, or just plain derping out, DerpGPT is here to entertain.

Python PyTorch
View Project
Build your own AI model from scratch project

Build your own AI model from scratch

AI model built based on the transformers architecture. Made in Python with PyTorch. The code is also explained on my youtube channel, where I made a series that guides the user through the project.

Python PyTorch
View Project View Youtube Guide