• 
  Design, build, and evaluate LLM agents that operate against real dealership systems — booking service appointments, answering vehicle availability questions, resolving customer identity, and escalating to humans with full context.
• 
  Own the tool-use layer: define the tools and function schemas agents call, the guardrails around each, and how the system fails when a downstream service is slow or unavailable.
• 
  Build agent evaluation infrastructure — offline eval sets, adversarial and edge-case suites, live A/B testing, and regression gates that block deployment on quality drops.
• 
  Design and implement escalation logic and confidence thresholds: where the agent acts, where it confirms, and where it hands off to a person.
• 
  Develop and maintain RAG systems over dealership content (service history, OEM documentation, policy, inventory) using Bedrock embeddings, pgvector on Aurora, and OpenSearch Serverless.
• 
  Own prompt architecture, versioning, and change control as a first-class engineering artifact under source control.
• 
  Build, validate, and deploy predictive models on lakehouse data — gross profit forecasting, customer lifetime value, defection risk, next-service prediction, identity resolution, and inventory pricing signals.
• 
  Own the full model lifecycle: feature engineering, training, validation, deployment, monitoring, and retraining. Ship models with drift and degradation monitoring from day one.
• 
  Convert business questions from operations and ownership into well-posed modeling problems and push back when a question is better answered with a query than a model.
• 
  Quantify and communicate model impact in dealership terms: gross, units, retention, CSI, labor hours saved.

Requirements

• 
  4+ years applying data science in production, with models that made real decisions and had real consequences.
• 
  Strong Python and SQL. You write code others can run and maintain.
• 
  Hands-on experience building LLM agents with tool use and function calling — not just prompt engineering. Be ready to walk through a system you built and how you evaluated it.
• 
  Practical RAG experience: embeddings, vector search, chunking, retrieval evaluation, and re-ranking.
• 
  Sound statistical fundamentals and honest handling of uncertainty. We prefer a well-calibrated interval over a confident point estimate.
• 
  Experience deploying models to production — not handing notebooks off to an engineering team.
• 
  Working comfort with AWS ML tooling (Bedrock, SageMaker, Lambda) and a lakehouse or data warehouse environment.

• 
  Databricks, Spark, and Delta Lake experience.
• 
  Experience with agent frameworks, orchestration patterns, and structured output / JSON-mode reliability.
• 
  Voice AI or conversational systems experience (Retell, Vapi, or comparable), including latency constrained design.
• 
  Time-series forecasting and causal inference exposure.
• 
  Automotive retail domain knowledge — DMS, CRM, F&I, fixed operations. A candidate who understands what an RO or a chargeback is will ramp faster.
• 
  Familiarity with AI safety and evaluation practice, including handling of PII in prompts and logs.

Benefits

• 
  Fully remote – work from anywhere in the Philippines.
• 
  Work on live agentic AI systems — not POCs, not slideware, not handoff-to-engineering.
• 
  Data layer already in place (AWS lakehouse, Databricks) so you can focus on modeling and agents, not plumbing.
• 
  Small, senior, high-autonomy team with documentation-first culture.
• 
  Opportunity to define the evaluation and deployment standards every new model and agent will follow.
• 
  A team environment that values intellectual honesty, technical depth, and follow-through.