AI TRAINER & LLM EVALUATOR
14 months of freelance work at RWS, reviewing prompts, model claims, conversation context, and transcriptions.
FLORIANÓPOLIS, BRAZIL
SELECTED WORK / FIVE SIMULATIONS
Fictional inputs. Each case shows the evidence and the reasoning behind the review.
CONTEXT RETENTION
The user originally wanted a short, informal message to Marina confirming Thursday at 2 p.m., with no emoji.
“It’s moved to Friday at 10 a.m. Keep everything else the same.”
“Hey, Marina! Confirming our meeting for Thursday at 2 p.m. See you then!”
The tone and length fit. The time is stale.
Use the corrected time. Keep the recipient, informal tone, and no-emoji instruction.
FACTUAL ACCURACY
The customer bought the item 12 days ago. The policy allows requests within 14 calendar days. Approved refunds are usually processed within five business days.
“Yes, your refund is guaranteed and will be in your account in five calendar days.”
The customer can still submit a request. The policy makes approval conditional, measures processing in business days, and gives no date for the deposit.
“Yes, you can still request a refund. If it’s approved, processing usually takes up to five business days.”
PROMPT RANKING
“Summarize the support conversation and explain what happened.”
“Summarize the conversation in up to three bullets for the billing team. Include the issue, what support checked, and the next action. Mark missing information as ‘not provided.’”
Prefer B. It identifies who will use the summary and what they need. If support never checked the charge, B gives the model a way to mark the gap without filling it in.
TRANSCRIPTION REVIEW
“I sent it on Tuesday… I mean, on Wednesday, to [inaudible].”
“I sent it on Tuesday to the finance team.”
The brief calls for preserving self-corrections and marking unclear speech. Restore “I mean, on Wednesday” and keep “[inaudible]”. The reference never establishes the recipient.
REVIEWER DISAGREEMENT
A shipping label was created. There is no carrier scan or delivery estimate.
A — “Yes, it will arrive tomorrow. A tracking number has already been created.”
B — “The shipping label has been created. There’s no delivery estimate yet.”
A answers directly, but B is the better choice when factual accuracy takes priority. Use B as the reference here and test the rule on an order where the carrier actually provides an estimate. If the rubric leaves the priority unclear, ask the project lead to resolve it before adding the case to the reference set.
PROFESSIONAL EXPERIENCE
14 months at RWS
Ranked prompts, checked claims in model responses, reviewed multi-turn conversations for context retention, and transcribed material for AI training.
Brazilian Portuguese · Native
English · Fluent
Prompt ranking
Context review
Transcription
User update
Model reply
Review
Mock speech reference
Resume PDF ↗
Full portfolio PDF
lorensanth@gmail.com ↗