Architecture
LocoPuente is the convergence point for the entire LocoLabo research programme. Five sibling projects each contribute a layer to what users experience.
Three Layers, Five Projects
Section titled “Three Layers, Five Projects”| Layer | Projects | What They Contribute |
|---|---|---|
| Infrastructure | LocoBanco + LocoConvoy | Hardware selection, capacity planning, multi-GPU scaling |
| Intelligence | LocoLLM + LocoAgente | Routed specialist models, agentic tools, scaffolded reasoning |
| Experience | LocoEnsayo | AI-populated rehearsal environments, simulation scenarios |
How Each Project Feeds LocoPuente
Section titled “How Each Project Feeds LocoPuente”LocoBanco feeds the infrastructure decisions
Section titled “LocoBanco feeds the infrastructure decisions”LocoBanco maps the floor of local LLM inference across every consumer GPU VRAM tier. That research answers the question every deployment starts with: what is the minimum viable hardware for a production service? LocoPuente PoC deployments are essentially LocoBanco findings applied to a real environment, validating benchmarks under genuine production load.
LocoConvoy handles concurrent load
Section titled “LocoConvoy handles concurrent load”When fifty users hit the service simultaneously, single-GPU inference becomes a bottleneck. LocoConvoy’s research into multi-GPU parallelism on consumer PCIe hardware, load balancing, and Mixture of Agents architectures feeds directly into LocoPuente’s capacity planning and scaling decisions.
LocoLLM provides the intelligence layer
Section titled “LocoLLM provides the intelligence layer”LocoLLM routes requests across a swarm of small specialist models behind a smart router, which delivers more efficient, more targeted responses at lower VRAM cost than one large model. As LocoPuente scales, LocoLLM is the architecture that makes it affordable.
LocoAgente powers the sophisticated tools
Section titled “LocoAgente powers the sophisticated tools”DeepTutor, the Curriculum Explainer, and the research assistant that can read a PDF and generate a structured summary all need more than single-turn chat. LocoAgente’s research into agentic scaffolding on small models is what makes these tools viable without frontier hardware.
LocoEnsayo supplies the experiences
Section titled “LocoEnsayo supplies the experiences”LocoEnsayo builds AI-populated rehearsal environments for professional education. In LocoPuente, those environments are the most substantial user-facing experiences: simulated organisations, role-play scenarios, virtual client meetings. The research and the deployment are the same system.
API-First Design
Section titled “API-First Design”All core AI services expose OpenAI-compatible APIs, so components can interoperate across the stack.
| Endpoint | Provided By | Consumed By |
|---|---|---|
/v1/chat/completions | Ollama | Open WebUI, Perplexica, AnythingLLM, Open Notebook, Custom chat |
/v1/audio/transcriptions | Speaches | Open WebUI, Open Notebook |
/v1/audio/speech | Speaches | Open WebUI, Open Notebook |
| Image generation API | ComfyUI | Open WebUI (in-chat images), direct UI |
| Web search API | SearXNG | Open WebUI, Perplexica |
This means any component can be swapped or upgraded independently without breaking integrations.