Данило

DANIEL BESAHA
AI Engineer | RAG, Agentic AI & Machine Learning
Brașov, Romania | GMT+3 | LinkedIn | WhatsApp | [відкрити контакти](див. вище в блоці «контактна інформація»)

PROFILE
AI Engineer with 3.2+ years of experience building production LLM applications, RAG platforms, agentic workflows, and machine-learning
systems. Hands-on with LangChain, LangGraph, LlamaIndex, LLM APIs, tool calling, MCP, multi-agent orchestration, hybrid retrieval, reranking,
evaluation, observability, and FastAPI/Docker integration. Experience spans retrieval and data pipelines, AI agents, production reliability and
optimization, classical ML, and computer vision.
EXPERIENCE
Nextcode.Tech Apr 2025 - Sep 2026
AI Engineer
Client Document Intelligence & Agentic RAG Platform - client system for reliable search, reasoning, and source-grounded answers
across internal business documents.
Owned major parts of a production RAG system across document ingestion, retrieval, evaluation, API integration, and production
improvements for a client project.
Improved retrieval Recall@5 from 72% to 85% by combining dense embeddings with BM25, metadata filtering, version-aware retrieval, and
cross-encoder reranking.
Reduced unsupported LLM answers by about 40% through relevance thresholds, evidence checks, citation requirements, and explicit no-
answer logic.
Built agentic and multi-agent workflows with LangChain/LangGraph including query routing and rewriting, specialized sub-agents, tool
calling, supervisor-worker patterns, checkpoints, fallback logic, controlled handoffs, and reusable skills.
Integrated AI components with FastAPI and Docker and used tracing and evaluation results to debug retrieval quality, latency, and failure
modes before production handoff.
TRIAGO Engineering | Brașov, Romania Aug 2023 - Mar 2025
Machine Learning Engineer
Lead Generation & Sales Intelligence Platform - internal ML system for discovering and prioritizing B2B manufacturers likely to need
industrial equipment.
Built LangChain-based LLM enrichment pipelines orchestrating source parsing, entity extraction, normalization, validation, retry/fallback
logic, and evidence aggregation before downstream ML scoring.
Designed a two-stage lead-ranking architecture with a lightweight candidate gate followed by richer profile-level scoring, reducing expensive
enrichment by about 44% while retaining around 92% of relevant leads.
Improved lead detection recall from 61% to 73% at about 70% precision by combining profile-level text, metadata, activity, and visual signals
with TF-IDF, embeddings, Logistic Regression, CatBoost, and feature engineering.
Contributed to computer-vision model development, testing image embeddings, aggregation methods, and feature sets for profile-level
lead scoring.
SKILLS
AI / LLM & Agentic Systems: LangChain, LangGraph, LlamaIndex, LLM APIs, agentic workflows, multi-agent orchestration, sub-agents, supervisor-
worker patterns, tool/function calling, MCP, checkpoints, handoffs, reusable skills, structured outputs, prompt engineering
RAG, Search & Evaluation: Qdrant, vector databases, embeddings, BM25, hybrid retrieval, metadata filtering, query routing/rewriting, cross-encoder
reranking, retrieval metrics, RAG evaluation, grounded generation, citations, no-answer/guardrail logic
Machine Learning & CV: scikit-learn, CatBoost, XGBoost, LightGBM, PyTorch, timm, TF-IDF, classification, ranking, clustering, PCA, feature
engineering, embedding-based CV features
Backend & MLOps: Python, SQL, PostgreSQL, FastAPI, REST APIs, Pydantic, Docker, GitHub Actions, MLflow, DVC, Git, Pytest, AWS
LLMOps & Observability: Langfuse, LangSmith, tracing, offline evaluation, error analysis, latency/quality monitoring, fallback and reliability patterns
EDUCATION & LANGUAGES
Priazovskiy State Technical University, Dnipro - Bachelor's degree, Automation and Computer Technology, 2021 - 2025. Degree officially
recognized in Romania.
Languages: English - B2 | Ukrainian - Native | Russian - Native