About this role
Summary: Design and build production-grade, multi-tenant AI pipelines using Python and LLM orchestration. Ensure scalability, observability, cost-efficiency, and security while enabling configuration-driven onboarding for enterprise customers.
You'll make Volta's AI pipeline production-grade at enterprise scale: multi-tenant, observable, cost-efficient and configurable per customer without code changes. You'll take parsing, matching and LLM orchestration services that already work and make them work for fifteen enterprise customers at once. You report to the Head of AI.
- Turn the Core into configuration: design the tenant configuration model (prompts, schemas, matching rules, thresholds, feature switches) so onboarding a new enterprise customer is configuration plus evaluation, not a code change
- Make multi-tenancy real: per-tenant data separation across services, storage, caches, knowledge bases and traces
- Scale the pipeline: Cloud Run services for ingestion, API, parsing and matching that absorb bursts of orders and large documents, with queues, idempotency and retries that don't re-run the whole pipeline
- Make it observable: every LLM call traced and costed in Langfuse, trace IDs carried onto orders, alerts on stuck orders and quality regressions
- Keep cost per order under control: model choice per step, caching, batching, cost dashboards per customer
- Ship safely: replay and regression tests in CI for prompts and schemas, versioned prompts and configs, gradual rollout per tenant
- Make accuracy a release gate: plug per-customer evaluation harnesses into the delivery pipeline, with the data scientists
- Make it secure: treat prompts, model outputs, tools and data flows as untrusted attack surface, and enforce it with automated guardrails, least-privilege access and continuous red-teaming in CI/CD
### Profile
- Backend engineer who moved into AI systems: 4+ years building production backend services in Python, at least one or two on LLM-powered systems in production
- Distributed systems done right: queues, async processing, idempotency, retries, back-pressure, concurrency on shared databases
- Multi-tenant SaaS: you've designed or operated a system serving many customers with real isolation requirements
- Observability first: you've instrumented services with tracing and metrics and used them to debug production
- Config over code: you've built systems driven by versioned configuration, feature flags or rules, and kept them understandable
- Must have: strong Python (FastAPI or similar), typing, tests; production LLM orchestration (structured outputs, tool calling, retries and fallbacks, prompt versioning); containerised services on Cloud Run, GKE, EKS or ECS; tracing and LLM observability (Langfuse, OpenTelemetry or similar); CI/CD and automated testing of non-deterministic systems; fluent English
- Nice to have: GCP (Cloud Run, Vertex AI, Pub/Sub, IAM); Elasticsearch or hybrid search; agent frameworks; cost optimisation of LLM workloads; GDPR and data residency in practice; French or Italian
- Location: Paris or Milan, hybrid, 3 office days a week
### Recruitment process
- Screen with the Chief of Staff (30 min): scope, seniority, motivation
- Technical deep dive with the CTO and the Head of AI (60 min): a simplified version of our pipeline: make it multi-tenant and configurable per customer, with cost and observability in mind. No take-home
- Culture and founder conversation (45 min)
- Reference checks