You will get LLM Evaluation Audit with Golden Set and Detailed Metrics Report
Rising Talent
Rising Talent
Project details
You will get a complete LLM evaluation audit that tells you exactly where your system breaks, what to fix first, and how much each fix matters.
My background. I designed and shipped AJUDEM, a production multi-agent RAG system at Petrobras (Brazil's largest energy company). Per-case review time dropped from several minutes to about 20 seconds, and monthly divergence cases fell from 10-15 to fewer than 3. For my undergraduate thesis I built a multi-agent essay-grading harness with conditional arbitration, validated through a controlled paired study with maximum human-vs-AI divergence of 1.22 across 24 cases.
The Standard tier includes a 30 to 50 case golden test set built from your domain, accuracy and latency metrics measured under controlled conditions, and a prioritized roadmap with effort estimates per fix. The Advanced tier extends with implementation of the top 3 fixes and 30 days of post-delivery support.
Stack. Python, LangGraph, LangChain, ChromaDB, FastAPI, Claude API, OpenAI API. Daily Claude Code user with ~19 MCP server stack.
Describe what you need evaluated. I'll respond with a specific plan.
My background. I designed and shipped AJUDEM, a production multi-agent RAG system at Petrobras (Brazil's largest energy company). Per-case review time dropped from several minutes to about 20 seconds, and monthly divergence cases fell from 10-15 to fewer than 3. For my undergraduate thesis I built a multi-agent essay-grading harness with conditional arbitration, validated through a controlled paired study with maximum human-vs-AI divergence of 1.22 across 24 cases.
The Standard tier includes a 30 to 50 case golden test set built from your domain, accuracy and latency metrics measured under controlled conditions, and a prioritized roadmap with effort estimates per fix. The Advanced tier extends with implementation of the top 3 fixes and 30 days of post-delivery support.
Stack. Python, LangGraph, LangChain, ChromaDB, FastAPI, Claude API, OpenAI API. Daily Claude Code user with ~19 MCP server stack.
Describe what you need evaluated. I'll respond with a specific plan.
AI Algorithms
Large Language Model, Transformer ModelAI Applications
AI Chatbot, AI-Enhanced Classification, Conversational AI, Natural Language Generation, Natural Language UnderstandingAI Development Language
PythonAI Tools
GitHub Copilot, Hugging Face, StreamlitAI Models
BERT, ChatGPT, GPT-3, GPT-4What's included
| Service Tiers |
Starter
$500
|
Standard
$1,500
|
Advanced
$3,000
|
|---|---|---|---|
| Delivery Time | 7 days | 14 days | 28 days |
Number of Revisions | 1 | 2 | 3 |
AI Model Integration | - | ||
Batch Normalization | - | - | - |
Database Integration | - | - | |
Detailed Code Comments | - | ||
Image Upscaling | - | - | - |
MLOps | - | - | |
Model Deployment | - | - | |
Model Documentation | |||
Model Monitoring | - | ||
Model Testing & Optimization | - | ||
Model Tuning | - | - | |
Natural Language Processing | |||
NLP Tokenization | - | - | - |
Pre-Training | - | - | - |
Prompt Engineering | |||
Setup File | - | ||
Source Code | - |
Optional add-ons
You can add these on the next page.
Extra Golden Set Cases (+50)
(+ 5 Days)
+$400
Live Walkthrough Call (60 min)
(+ 1 Day)
+$150
Detailed Documentation Pack
(+ 3 Days)
+$250Frequently asked questions
About Lucas Lorenzo
AI Engineer | Multi-Agent, RAG & LLM Evaluation | Python + LangGraph
Rio de Janeiro, Brazil - 10:11 pm local time
At Petrobras (Brazil's largest energy company), I designed and shipped AJUDEM, an internal multi-agent RAG system that reviews monthly stock-divergence cases from a partner company. The system retrieves over 100 validated historical cases, scores new cases on a 0 to 10 rubric with reasoning, and routes them through human-in-the-loop validation. Production outcomes: per-case review time dropped from several minutes to about 20 seconds, and monthly divergence cases fell from 10-15 to fewer than 3. I explored DSPy for prompt optimization on the AJUDEM pipeline; data volume in the corpus was the limiting factor for full deployment.
For my undergraduate thesis I built an essay-grading harness with conditional arbitration, inspired by the official process Brazil uses for its national exam (ENEM). Two independent grader agents evaluate each response in parallel, and a third arbiter agent is invoked only when their divergence exceeds a configurable threshold. The system reached 100% end-to-end completeness across the test set, with a maximum human vs AI divergence of 1.22 across 24 graded essays. That was under the 2.0 escalation threshold, so the arbiter was never triggered. I also measured RAG impact by difficulty: +0.84 on intermediate cases, +0.12 on weak cases, and 0 on extreme cases.
I'm currently building a framework that compares LLM outputs from multiple perspectives and assembles them into structured quality assessments.
Stack. Python, LangGraph, LangChain, ChromaDB, PostgreSQL, FastAPI, Claude API (messages, tool use, prompt caching), OpenAI API, MCP servers. I work daily inside Claude Code with a ~19 MCP server stack, so I'm comfortable building agent systems with persistent memory, sub-agents, and custom skills.
Describe what you need evaluated, scored, or automated. I'll respond with a specific plan.
Steps for completing your project
After purchasing the project, send requirements so Lucas Lorenzo can start the project.
Delivery time starts when Lucas Lorenzo receives requirements from you.
Lucas Lorenzo works on your project following the steps below.
Revisions may occur after the delivery date.
Discovery and scope alignment
Kickoff call (45 min) to confirm scope, success criteria, and target metrics. Review system access and current architecture documentation.
Golden test set construction
Build 30 to 50 validated cases sampled across difficulty tiers (weak, intermediate, extreme). Include ground truth and acceptance criteria for each case.