Hi, I'm André Codea 👋

Jr. AI/ML Engineer | LLMOps Analyst | Cognitive Architect

Building the bridge between Deterministic Engineering and Probabilistic AI.

I am a Jr. AI Engineer and Computer Science student focused on architecting autonomous intelligent systems. My expertise lies in LLMOps and Prompt Engineering. I am driven by the tangible impact that my solutions generate and potentialize. In my free time, I like to play videogames, study, code, lift weights, and cook!

Currently, I'm focused on engineering Machine Learning Systems, Cognitive Systems, and preparing for the Microsoft Azure AI Fundamentals (AI-900) certification (wish me luck!).

Multi-agent research pipeline with semantic memory and knowledge curation via HITL.

• Architecture: Orchestrated an advanced StateGraph using LangGraph featuring four specialized agents (Orchestrator, Router, Researcher, Writer).
• Control: Implemented Human-in-the-Loop (HITL) breakpoints for knowledge base curation in ChromaDB, ensuring zero hallucinations before web searches via Tavily API.
• Observability: Full-node tracing and PromptOps via LangSmith, exposing reasoning traces through a mandatory <thought> protocol before any tool call.
• 🏷️ LangGraph LangSmith HITL ChromaDB PromptOps

Real-time market analysis API with Generative UI and aggressive payload compression.

• Optimization: Slashed token consumption by 60% (from ~100k to ~40k) via a payload compression pipeline (JSON → CSV + pruning).
• Telemetry: Actively monitored Time To First Token (TTFT) and latency via LangSmith.
• UX Innovation: Delivered real-time financial dashboards via Generative UI (Thesys SDK), orchestrating LangChain and Yahoo Finance with Chain-of-Thought reasoning.
• 🏷️ FastAPI LangChain GenUI LangSmith

Cognitive Operating System for personal productivity via Telegram, governed by strict guardrails.

• Memory: Three-tiered architecture inspired by the Atkinson-Shiffrin model (Sensory buffer in PostgreSQL, Short-term window, Daily consolidation via RAG in Supabase/pgvector) using Llama 3.3 70B.
• Security: Deployed guardrails using Llama 3.1 70B (0.7 threshold) for NSFW and jailbreak detection.
• Orchestration: Integrated specialized sub-agents (Calendar, Gmail, Tasks) via Model Context Protocol (MCP) with mandatory Chain-of-Thought reasoning.
• 🏷️ n8n MCP Supabase/pgvector Llama 3.1

Dashboard for Sales Intelligence, Customer Clustering, and AI-driven narrative reporting.

• Machine Learning: Built a customer segmentation pipeline via K-Means and a weekly revenue forecasting model using Prophet with 95% confidence intervals.
• Observability: Tracked LLM latency in real-time via token-by-token streaming (TTFT P50: ~800ms).
• Reporting: Generated narrative PDF reports combining ML trends with LLM analysis via ReportLab.
• 🏷️ Python Scikit-learn Prophet Streamlit

Let's build the future of autonomous systems. 🚀