About the team
The AI Automation team builds every AI-shaped tool the company needs as it scales. The team is small — three people — and each member owns a subset of services end to end. Today that covers chatbots, voice bots, document analysis, customer verification, and communications monitoring.
The role
Aviva's customers reach us through chat long before they reach a person: they ask about their balance, negotiate a payment, book a visit to a kiosk, renew a loan, or answer the questions of an onboarding interview. Those conversations are handled by a squad of LLM agents we build in-house on Google's Agent Development Kit, and they are already on the critical path of collections, growth, and origination.
You will work towards that platform's next stage: a general-purpose squad that keeps growing, voice agents for collections negotiations and marketing campaigns, a new hiring agent squad that recruits kiosk managers, and the evaluation harness that makes all of it measurable instead of anecdotal.
This is a hands-on role. You will write production code, design agent behaviour and conversation flows, and work directly with collections, growth, HR, and the engineering teams behind Aviva's core systems to turn what the business wants said into something an agent can safely say.
The system
The agent platform is a multi-service system that serves several live bots from one codebase. A coordinator agent routes each conversation to a specialised subagent — collections, growth, KYC, renewals — and every subagent works through deterministic tools that own the truth: loan balances, payment instructions, kiosk availability, appointment booking, handover to a human.
Tools never answer the customer directly. They return typed cues — guidance the model reads, transfers it must perform, or channel events (text, images, handovers) the runtime delivers — so behaviour stays inspectable and a hallucinated number can be caught before it is sent. Around that sit the parts that make it a product: RocketChat and Facebook Messenger channels, the APIs of Aviva's core lending systems, Postgres-backed sessions with Redis locks for concurrent turns,



a Celery worker that follows up on reminders, a model gateway where every call asks for a capability tier rather than a provider, per-turn tracing with LLM judges, and BigQuery + Dagster analytics behind the dashboards the business reads.
Streaming voice is already wired into the platform — one of our agents runs on the Gemini Live API over a WebSocket the gateway proxies, sharing the same tools and prompts as its text twin — so a new voice agent starts from a working integration rather than a blank page.
Key responsibilities
Agent development
Design and ship agents end to end: conversation design → prompts and instructions → typed tools → guardrails → evals → production rollout.
Build new subagents and new squads for new business flows, reusing the shared platform rather than forking it.
Integrate the systems agents act on — Aviva's core lending APIs, CRM, kiosk and scheduling data, messaging channels — as async clients with real timeouts and error isolation.
Design guardrails for the failure modes that matter in lending: invented amounts, invented policy, internal plumbing leaking into a customer message, promises we cannot keep.
Architecture
Keep the boundary between agent reasoning and deterministic tools sharp — the tool owns the fact, the model owns the wording.
Drive incremental refactors that let squads, channels, and the runtime evolve and deploy independently.
Help improve the current agent testing suite, currently composed by googel ADK evals and Langfuse evalautors
Quality and reliability
Maintain unit tests for tool logic and eval suites for conversation behaviour.
Monitor handover rates, guardrail triggers, tool failures, judge scores, token cost, and latency through traces and BigQuery dashboards.




Own the escalation path when a live agent misbehaves: customer-facing wrong answers are handled as incidents.
API design & integration
Aviva's core lending platform and the messaging channels are the main consumers of these services: schema and endpoint changes need versioning, contract tests, and coordination with the teams on the other side.
Be the point of contact between Data Science and the business owners of each bot — collections, growth, HR — translating what they need said into agent behaviour, and explaining honestly what an agent can and cannot be trusted to do.
Benefits
Attractive compensation package.
Fast-paced environment with significant growth opportunities.
15 annual vacation days + 7 annual personal days.
Option to work remotely 3-4 days per week ; or fully-remote (as long as you can come to CDMX ~twice a year)
Adaptable work schedule
Requirements:
Required skills
5+ years of Python in production systems.
LLMs in production, not just prototypes: prompting, tool and function calling, an agent framework of some kind, LLM evaluation experience.
Containerisation and orchestration — Docker and Kubernetes in production: resource limits, rolling deploys, health probes.
Experience owning production services consumed by other teams.
Judgement about conversation design: distinguishing a reply that is technically correct from one a customer under financial stress will accept.
Communication. You will be the main point of contact between Data Science and other teams for a subset of services; precise communication matters as much as precise code.
Fluency in Spanish and English. The agents speak Mexican Spanish to customers and most meetings are in Spanish; code and documentation are in English.
Nice-to-haves
Google ADK, Vertex AI, or another agent framework used in production.
LLM observability and evaluation tooling (Langfuse, LangSmith, OpenTelemetry, DeepEval).
Dagster and BigQuery for analytics pipelines and dashboards.
The Azure ecosystem (AKS, ACR, Azure DevOps) or an LLM gateway such as TrueFoundry.
RAG and equivalent LLM memory techniques

📌 Llm Engineer — Agentic Scpecialist (Xico)
🏢 Aviva
📍 Xico