Organics R&D AI

Organics’ cosmetics R&D team had its AI help split across separate assistants — one for stock and raw materials, another for sales formulation — each with its own history, while ingredient references, current stock and past formulas lived in different places. Checking whether a formula could be made from materials in stock, and at what cost, meant moving between tools. The brief: one agent that can search materials, confirm stock, draft and cost a formula and research the market in a single conversation, while making clear what is confirmed, what is an estimate and what still needs the lab.

We replaced several disconnected AI paths with one ReAct agent whose behaviour is written down: each skill is a versioned Markdown card that names the tools it may use and the rules it has to follow.

1. 1. One ReAct agent on Gemini

   Every request runs Plan → Act → Verify → Synthesize. The screen shows the high-level plan and the tools used, never the model’s private chain of thought.

2. 2. Skill cards loaded at runtime

   The five Markdown cards are read at runtime and injected into the agent’s instructions, with a compact fallback so an incomplete deployment never answers from half the guidance.

3. 3. Retrieval over the team’s own data

   Qdrant semantic search covers five collections, exact MongoDB queries handle codes and fields, a stock tool always reads the current stock source first, and grounded web search returns its source URLs.

4. 4. Formula tools with a version log

   The agent can generate, find, load with comments, revise and confirm formulas; each confirmation bumps the version in a log, and costing and scaling run through a calculator that labels estimates.

5. 5. Evaluation and a closed perimeter

   A RAGAS and Promptfoo harness scores faithfulness, context precision and recall, and latency, and includes a prompt-injection suite. Protected API routes require a session, and the acting user always comes from that session, never from the request body.

R&D AI runs in production as one workspace for the team. The old assistant links redirect into it, and earlier stock and sales conversations carry on in the same history.

• The whole question in one chat

  Formulators and sales staff find a material, check it is in stock and cost a draft formula without switching systems.

• Clear about what is confirmed

  Confirmed stock is kept apart from catalogue references, estimates are labelled, and lab validation is always called out.

• Formulas with a history

  Every formula keeps its versions and comments, and nothing the AI drafts becomes final until someone confirms it.

• Market answers with sources

  Research questions come back with cited links, and the agent says so when current information could not be verified.