Experience

Backend engineer at Publicis Sapient.

I work on agents and orchestration for a platform that generates insights on demand from whatever data a user connects.

Agent orchestration & harness

  • Built the orchestration layer routing user questions across an NL2SQL engine, a RAG pipeline, and structured data connectors.
  • Designed the routing logic that decides which path a question takes — and degrades gracefully when it belongs to none of them.
  • Handled the unglamorous half of agent systems: retries, partial failures, timeouts, and keeping state coherent across multi-step queries.
  • Users can register their own MCP servers; the agent routes to them alongside the native paths, which meant designing for tools we don't control.
  • Handled multiple response types: a question might resolve to a number, a table, a KPI, or a chart, and the harness decides which.
  • Building a centralized, version-controlled prompt registry and the evals that keep it honest across versions.

Natural language to SQL

  • Worked on an NL2SQL engine generating and executing queries against live user databases.
  • Handled dialect differences across Postgres, Databricks, BigQuery, and Salesforce Data Cloud. The same question compiles differently, and sometimes not at all.
  • Built guardrails around generated SQL so a malformed question can't become a destructive query.

Retrieval over unstructured data

  • Built the retrieval pipeline for text-based sources alongside the structured-data path.
  • Worked on chunking, retrieval quality, and grounding responses in the source material.

Latency: gRPC query layer

  • Moved query execution to gRPC. Protobuf replaced JSON serialization of SQL result sets, and persistent HTTP/2 connections removed the per-request handshake.
  • For large result sets, gRPC server-streaming replaced HTTP batch-and-poll — rows flow back as they're read instead of being materialized and paginated per request.

Data connectors

  • Built ingestion across Postgres, Databricks, BigQuery, and spreadsheet uploads.
  • Normalized schema handling so downstream agents see a consistent surface regardless of source.
  • Built robust data dictionaries as an ontology store for column semantics, business meaning, and relationships, decoupled from the raw schema. The offline async pipeline runs across webhook-coordinated microservices.