Claude Production Review
Audit of an existing agent or prototype: architecture, security, tenant isolation, cost, evals; written plan.
Nicolas Trossat
Claude Certified Architect — Foundations
Independent Claude Certified Architect. US LLC, Eastern hours.
eCommerce platforms, multi-tenant B2B systems, analytics platforms, and now AI at scale.
20+ yearsMulti-agent orchestration, MCP servers and document extraction with structured responses, on the Claude API and AWS Bedrock.
Claude · MCPFastAPI and Pydantic services, dbt semantic layers, Lightdash BI on Databricks and Redshift, SSO.
FastAPI · dbtClaude Certified Architect, by Anthropic. MS in Engineering, Arts et Métiers ParisTech.
CertifiedAudit of an existing agent or prototype: architecture, security, tenant isolation, cost, evals; written plan.
One agent or MCP server built to production on your stack.
Technical direction, code review, Claude Code adoption.
“Nicolas asks the right questions: he takes the time to understand the problem deeply before jumping to a solution.”
I own a project from requirements through production: design, build, code review and rollout.
Agents, MCP servers and APIs, with SSO, tenant isolation and human validation handled in production.
Technical direction, code review, and Claude Code adoption for existing engineering teams.
Internal agent platform exposing agents to every other service in the company.
ImpactThe company's software and services gained new AI features and agents.
ComplexityMulti-agent orchestration behind a FastAPI gateway. SSO, tenant isolation, security and enterprise integration.
StackClaude, AWS Bedrock, MCP, FastAPI, SSO.
# No raw SQL: the agent resolves the question # into a governed dbt Semantic Layer query query = { "metrics": ["revenue"], "group_by": ["customer__region"], "where": "metric_time >= '2026-04-01'" } # MetricFlow compiles the SQL. One definition drives # the schema, the dashboards, the API and the answer.
Multimodal agent extracting metadata from documents, with the fields specified per request and returned as a structured response.
ImpactFields extracted on demand. Low-confidence results go to a human for validation.
ComplexityMultimodal input, per-request field specs, structured response, confidence scoring, human-in-the-loop routing.
StackPython, Pydantic, Claude, AWS Bedrock.
The full analytics stack of a multi-tenant platform: self-service BI on Lightdash, APIs and a text-to-SQL agent, on Databricks and Redshift.
ImpactReporting moved out of legacy software. Business teams and clients reach their own data and dashboards directly.
ComplexityEntra ID SSO, row-level filtering for tenant isolation, dbt pipeline and semantic models, analytics content managed as code.
StackPython, FastAPI, dbt, Lightdash, Databricks, Redshift, Entra ID.
Short and long-term projects, remote from Andorra.
US LLC, invoiced in USD, working on US Eastern Time.