Daniel Berhane Araya

About

AI/LLM engineer with published work who built FinVet, a production financial misinformation detection system using multi-pipeline RAG, confidence-based voting, and evidence-backed outputs (F1 0.85, +37% vs baseline). Also built NexusForest-MCP, a Dockerized Model Context Protocol server exposing structured global deforestation/carbon data via SQL tools for reliable LLM tool use. Previously delivered borrower risk-rating (PD) models at BMO Financial Group that were validated and integrated into an enterprise credit system through close collaboration with credit officers and portfolio managers.

Experience

Education

Key Strengths

• Built and deployed production AI systems with multi-component architectures (RAG + APIs + UI)
• Designed financial misinformation detection with evidence, sources, and confidence scoring
• Improved model performance to F1=0.85 (~37% over baseline) and published the work
• Reliability-focused design: safe failure modes (e.g., 'not enough information') and strict schema validation
• Strong evaluation discipline: curated test suites (100 queries), edge-case coverage, and retrieval metrics (precision/recall)
• Effective orchestration of agentic workflows using LangGraph state-machine/graph patterns
• Integrated LLMs with structured climate datasets via MCP + SQL tools; solved sparse/wide-table issues through normalization
• Cross-functional delivery: translated credit risk models into business logic for non-technical stakeholders; enterprise adoption after validation