ProtForge is an independent AI engineering practice in Barcelona. We measure where LLM products hallucinate, fix the pipeline behind it, and keep inference costs honest. Fixed-price audits, fractional architecture, no retainers you don't need.
Book a reliability audit See the three servicesWe only take work in fields where we have built the systems ourselves. Each practice has its own buyers and its own proof; all three share the same rule, that nothing ships without a measurement behind it.
For biotech, pharma and research teams
Generative and predictive models for proteins, antibodies and biosignals: model design and review, training and evaluation pipelines, and independent scientific due diligence of AI claims before a partnership or an investment.
For companies shipping LLM products
Where does the model make things up, how often, and what does it cost? Groundedness and hallucination evaluation built from your own traffic, regression suites in CI, and the routing and cost-per-task decisions that follow from the numbers.
For development and design agencies
You sold the AI feature; we build it under your brand. Retrieval assistants, document extraction, agent workflows and the evaluation suite that proves each one works, delivered by a senior engineer on a fixed scope per feature or one to two days a week. NDA as standard, your client relationship and margin stay yours.
Other LLM systems work is taken on request when it fits one of the three.
Everything we do starts from the same question: can every output be traced back to something true? We build the measurement first, then fix what it finds.
Two weeks on your production system. We build a groundedness and hallucination evaluation set from your real traffic, run it against your current pipeline, and hand back a ranked failure taxonomy with the fixes that matter.
Advisory for teams whose model bill or latency has stopped making sense. Model routing, caching, batching and evaluation gates in CI, with cost per task as the number everyone watches.
One or two days a week inside your team. Agent runtimes, retrieval pipelines, evaluation layers and the production habits that keep them honest. For companies that need senior ownership before they can hire it.
Products we built end to end, from first commit to paying users. Client engagements under NDA are described in a call.
A workers' compensation class-code checker for US businesses: a focused LLM-assisted lookup where the answer always points back to the classification source it came from.
WhatsApp ordering for restaurants, built solo from idea to production: conversational ordering flow, restaurant dashboard, digital menus and multi-tenant auth.
EU AI Act assessment for companies shipping AI systems: guided risk classification, LLM-assisted obligation analysis and automatically generated compliance documentation.
ProtForge started in 2023 as a generative-AI company for therapeutic antibody design: diffusion models, graph neural networks and protein language models on AWS. The science taught us one habit we never dropped: a prediction nobody can verify is worth nothing.
Today ProtForge is an independent engineering and advisory practice led by Guillermo Llopis, a two-time startup CTO and PhD researcher in LLM hallucination detection. He has shipped production multi-agent LLM platforms where every figure in every output traces back to a source document, and he brings that same evaluation discipline to every client system.
We build on Vercel, Supabase, Cloudflare Workers, AWS and the Anthropic and OpenAI APIs, and we are vendor-neutral when it comes to which model you should actually run.
Send a short note about your product and what you suspect is failing. You will get a reply from an engineer within one business day, and a scoped proposal within three.
guillermollopis@protforge.comWe take on a small number of engagements at a time so each one gets senior attention.