AI Recipe Authoring Agent

Overview

A pharmaceutical manufacturing execution (MES) environment depends on recipes — heavily regulated process definitions that govern how drugs are produced on the shop floor. Authoring them is slow, expert-bound work: an engineer translates an approved manufacturing recipe into a structured system configuration, validates it against site standards, and iterates with reviewers. This platform uses an agentic AI workflow to draft, validate, and refine those recipes with a human-in-the-loop (HITL) approval gate at every critical step.

The architecture below is generalized. Company-specific system names and internal URLs are intentionally omitted.

The engineering problem

Recipe authoring took ~4 months per recipe. The bottleneck was not typing — it was assembling knowledge: historical recipes, site-specific rules, global standards, and validation requirements scattered across enterprise systems. A pure “LLM writes the recipe” approach fails here for three reasons: regulated outputs demand auditability, model

Architecture

Technical design

• Agent layer — a stateful LangGraph workflow (draft → validate → retrieve → review → refine) where every transition is a persisted checkpoint. Long-running authoring sessions survive restarts and can be resumed or audited.
• Models — Amazon Bedrock for model access. Model choice is a configuration concern, not a code concern: prompts are versioned and evaluated independently of deployments.
• Retrieval — MCP tool servers expose a three-tier retrieval strategy: site-specific knowledge first, then global standards, then historical recipes — with pgvector similarity search over vectorized enterprise configurations (fed by a dedicated ETL pipeline).
• HITL gates — domain experts approve or reject agent output at defined checkpoints. Every decision is captured as workflow state, giving a complete audit trail.
• Observability — LangSmith traces every run; prompt and model changes ship through an evaluation harness before touching production flows.

Key engineering decisions

• Why agentic, not a single prompt? The task has real control flow: validation failures must loop back into drafting with new context. A graph with typed state beats prompt-fu.
• Why MCP for tools? Tool interfaces outlive agent frameworks. A standards-based tool layer let agents — and later, other AI systems — reuse the same retrieval surface.
• Why pgvector over a dedicated vector DB? Configurations already lived in PostgreSQL; keeping vectors beside the relational source of truth removed a sync problem and met enterprise operational requirements (backup, access control, HA) out of the box.
• Why HITL instead of full automation? In a regulated domain, trust is earned incrementally. The HITL gate is what made adoption possible — experts only accept what they can inspect.

Challenges & solutions

• Retrieval precision across sites — site-specific rules sometimes conflicted with global standards. Solved with the tiered retrieval strategy and explicit precedence rules.
• Non-determinism in regulated output — structured output schemas plus validation nodes in the graph; anything failing schema or business rules never reaches a reviewer.
• Evaluation drift — prompts changed frequently early on; versioned prompts plus LangSmith eval sets turned “it seems better” into a measurable regression gate.

Results

• Recipe authoring effort reduced from 4 months to 10 days.
• Expert review time became the primary constraint — the system shifted human effort from assembly to judgment.
• The MCP retrieval layer was subsequently reused by other AI platforms in the organization.

What I learned

The hard problems were not model problems. Retrieval quality, state modeling, and review ergonomics determined the outcome; the LLM was the easy part.

What I would change today

I would invest in automated evaluation before the first production deployment rather than after, and push more validation logic into deterministic tool calls instead of asking the model to self-check.

Technologies

LangGraph · Amazon Bedrock · MCP (FastMCP) · RAG · pgVector · Aurora PostgreSQL · FastAPI · React · LangSmith · Docker · ECS · GitHub Actions

knowledge goes stale, and hallucinated parameters are unacceptable in GxP manufacturing.

My role

I architected the platform end to end: the agentic graph design, the retrieval architecture, the MCP-based tool layer, the HITL checkpoint model, and the observability stack — and led the technical design reviews with senior leadership and site stakeholders.