Case — Oxiquímica (R&D Formulation Platform with Bayesian Optimization)

Type: AutoU (client: Oxiquímica — chemical/fertilizer industry) Role: Technical discovery, architecture and prototyping — translating a lab process into a software product Status: Live (in production) Stack (defined): Python 3.13, FastAPI, React 19 + TypeScript, Vertex AI (Google Cloud), BayBE (Bayesian optimization, Merck's open-source project), RAG with source governance, PostgreSQL

Confidentiality note: AutoU client project — validate what can be made public before exposing name/details.

Context and problem

Oxiquímica's R&D lab develops formulations (fertilizers and the like) by trial and error: each iteration goes through a ~14-day greenhouse/climate-chamber cycle, and the knowledge of which combinations work lives in the analysts' heads. Many iterations mean months until a stable formula.

Solution (designed product)

An R&D platform with an AI agent ("Colibri") that generates/assists formulas and reduces physical iterations:

Formulation Management Hub: KPIs (success rate, average iterations), cards per formulation with stability probability and a countdown for the climate test
Reactive composition table: autocomplete from the raw-materials database, auto-filled guaranteed content, suggested average purity (editable per batch) and real-time computed guarantee (concentration × purity) — computed on the front end and revalidated on the backend
Bayesian optimization with BayBE: suggests the next formulation to test based on previous experiments — fewer greenhouse cycles to reach the target
Scientific chat with dual RAG: answers by querying the internal knowledge base and web search with a source whitelist/blacklist (citation governance, every answer traceable to its origin)
Explicit formula versioning (V0, V1... with author and status) and a process timeline (definition → AI analysis → climate chamber → opinion → approval)

Technical work performed

In-depth study of BayBE (Merck's framework) with my own versioned Python examples — feasibility validated before promising anything to the client
Full architecture documented from the Figma prototype + designer comments, turning each comment into an executable specification (field-by-field behavior of the composition table)
Documented architecture decisions: reactive calculation on the front end with backend revalidation as the source of truth; purity override per composition without changing the raw-material record; default-block policy for unlisted web sources
Survey and cataloging of the client's material archive (general catalog and a detailed raw-materials map)
Observability with Grafana + Prometheus implemented by me: monitoring of cost (AI usage), consumption and platform infrastructure

Challenges and solutions

Complex scientific domain: formulation chemistry translated into a domain model (guarantees, purities per batch, raw-material functions) validated with the designer and the client
Responsible AI: in a confidential industrial environment, the web RAG has explicit source governance — recommendation to block by default
Prototype-driven discovery: open questions documented and addressed with the client instead of silent assumptions

Results and impact (expected)

Reduction in the number of physical iterations per formulation via Bayesian optimization [baseline: 4.2 average iterations in the prototype; target TO CONFIRM]
Lab knowledge captured in a queryable base (RAG) instead of tribal knowledge
Greenhouse cycle tracked digitally with informed interruption