gofri Why two g's? When I was a kid my grandma called me Gofri, and the family picked it up. She said the G hard, almost twice, so I write it twice.
production

Support-bot orchestration

for
Mercado Libre
when
2021–22
Problem
A support-bot prompt that was getting worse as it grew.
My contribution
Built the model integration and orchestration backend, routing structured LLM intents to internal services and integrating it across consuming teams. Informal tech lead alongside two senior developers.
Decision
Stop growing one prompt. Split it into retrieval, reclassification and answering; let the model ask for information in a structured form and the backend fetch it; hand off to a person when the bot can't help.
Delivered
Split into retrieval, reclassification and answering, with human escalation.
Result
About 25% of support handling steps were bot-handled, as reported by the manager and in performance reviews.
Status
production · 2021–22

≈25%of support handling steps

what I foundEvaluations showed the prompt regressing as it grew.

what I foundFailures included malformed structured output, internal information requests reaching customers and more hallucinations.

  1. You are a support assistant.
  2. Answer in the customer's language.
  3. If the order is late, explain the carrier states.
  4. If the customer asks for a refund, apply the policy.
  5. If the message is unclear, ask one question.
  6. If unsure, hand over to a person.

evals flag regressions as the prompt grows

  1. retrieval
  2. reclassification
  3. answering
human escalation
One prompt, split into stagesillustrative prompt lines

try it

Where a conversation ends up

Bot answers. The bot had what it needed and answered.

invented messages, illustrative routing

control

The model never called the services

  1. customer → interface

    1The customer writes a message

  2. interface → backend

    2The interface forwards it

  3. backend → model

    3The backend asks the model

  4. model → backend

    4The model returns a structured request for information

  5. backend

    5The backend fetches the internal information

  6. backend → model

    6The backend hands the information to the model

  7. backend → interface

    7The backend routes the answer back

  8. interface → customer

    8The answer reaches the customer

The backend is the controller: it reads the model's structured request, fetches the information and routes the response.

evaluations

What the evaluations flagged

  • Malformed structureThe model's structured output stopped being well formed.
  • Internal requests leakingInternal information requests reached customers.
  • More hallucinationsInvented answers became more frequent as the prompt grew.

The fix: split the prompt into retrieval, reclassification, answering, with human escalation.

counting

How a step is counted

One conversation, counted step by step

  1. Bot answers the first questionbot+1 = 1
  2. Bot answers a follow-upbot+1 = 2
  3. Bot reclassifies the conversationbot+1 = 3
  4. A person resolves ithuman+1 = 4

Total: 4 steps. Bot and human handling count alike.

invented conversation, illustrative

adoption

The shared interface, used widely

  • Mercado Libre
  • Pago
  • Shops
  • Envíos
  • other teams

ecosystem adoption, not a counted census

ownership

Who did what, and when

  1. evals show regressions
  2. prompt split in stages
  3. human escalation
  4. adopted across the ecosystem
  • MeOrchestration, model integration, API backend foundation and integrations into consuming teams; informal technical lead
  • Two senior developersAlongside me
  • ColleaguesThe shared microfrontend and additional ticket operations
  • Evaluation designWritten by others; I ran them

What shipped

Evaluations caught the regressions. The fix was structure, not a longer prompt.

  • Colleagues' shared customer-experience microfrontend was adopted across Mercado Libre, Pago, Shops, Envíos and other teams before I left.
Method and caveats

The 25% counts handling steps, not tickets, and carries no time-savings measure.

  • Models, evaluation dataset, traffic and rollout percentages: not available