Nicholas Hopper

AI Engineer, APX Stream

Nicholas Hopper is the AI engineer on APX Respond, an asset management SaaS product that automates due diligence questionnaire (DDQ) and request for proposal (RFP) responses with large language models.

Hopper joined APX Stream in June 2026. He designed the product’s agent architecture, its retrieval pipeline and the evaluation suite that every prompt or model change must pass before release.

What Nicholas Hopper did at APX Stream

  • Designed the DDQ and RFP auto-population engine in FastAPI and LangGraph: a multi-agent planner and executor with MCP tools and human approval gates.

  • Built retrieval end to end: OCR and layout-aware document parsing, chunking, pgvector hybrid search, reranking, a knowledge graph and grounded citations.

  • Built the evaluation suite that gates every prompt and model change on scored regression sets.

  • Owns inference: managed and self-hosted open-weight models behind a LiteLLM gateway with routing, fallback and cost caps, and fine-tuned models to improve answer accuracy.

  • Secured agents that handle confidential financial data with prompt-injection defenses, least-privilege tools, PII redaction and audit logging.

  • Wrote the specifications and review standards for prompts, evals and agent code, and advises leadership on build versus buy, model choice and cost.

  • Shipped the React 19 analyst workspace and deploys on Azure Container Apps with Bicep.

  • Started the document auto-population engine as a standalone prototype with a typed tool registry and per-tool benchmarks, then moved it into the product monorepo.

  • Routed all model calls through a self-hosted LiteLLM gateway so services hold no provider keys, with per-key spend tracking and model aliases as the swap point between providers.

  • Added a plan-then-draft gate so the model judges which retrieved sources are relevant before drafting, and refuses when none are.

  • Added an LLM-as-judge entailment check on drafted answers that fails closed when the judge is unavailable or its output cannot be parsed.

  • Moved prompts into versioned files that reload on every call, with the hash of the live text recorded in each call's provenance.

  • Built golden-set benches for document extraction tools, scored on recall, precision and exact-answer accuracy.

  • Built DOCX and XLSX parse and render codecs that write answers back into the original file layout.

  • Built the backend-for-frontend that composes the engine services behind one OpenAPI-first contract with a generated typed client.

  • Built a design system with a Storybook catalog and enforced it with lint bans and an exemption list that may only shrink.

  • Added distributed tracing across services and the model gateway with OpenTelemetry.

  • Wrote the Bicep infrastructure and Azure Pipelines for an Azure Container Apps environment with private PostgreSQL.

  • Wrote an end-to-end Playwright run that drives the real Docker stack twice from clean volumes.

  • Wrote agent skills for Claude Code and Codex covering service scaffolding, QA campaigns, bug registry upkeep, and delivery from parallel git worktrees to main.

  • Built the company's public marketing site in Astro with a markdown twin of every page and a generated llms.txt.

Skills used at APX Stream