AI Knowledge Bases
We connect your documents and databases to AI. Your teams ask questions in natural language and get accurate, sourced answers. Access controls match your business rules.
AI · Machine Learning · Data Engineering
We build AI systems for production.
Hoja designs and ships machine learning models, agents, and custom software. We turn your data into systems your team uses.
Services
What we do
From data pipelines to production AI products. Pick what you need, or let us scope the whole engagement.
AI Agents
We build agents that act on your behalf. They read data, act within limits you set, and connect to the CRM, messaging, and project tools your teams already use.
Data Engineering
We build pipelines that collect and serve your data. Your teams get reliable access to information. Your models stay accurate with fresh data.
Custom Software
We build web applications and APIs that fit your workflow, with clean architecture and tested code. Your team maintains the product after launch.
MLOps & Infrastructure
We ship AI systems into production and keep them running, with automated training, monitoring, and scaling. It keeps running without someone watching it.
Technical Strategy
We assess your data, pick the right approach, and scope the build. You get a plan with timelines and a clear yes or no.
Technologies
Our technology stack
We lean toward open-source models and vector search you can own and run on your own infrastructure. As a consultancy, we also work with the stack your team already runs.
Open weights
Qwen
Kimi
DeepSeek
Z.ai
Mistral
Hugging Face
Hosted APIs
OpenAI
Anthropic
Gemini
Grok
OpenRouter
Fireworks
Serving & orchestration
vLLM
Ollama
Ray
LangGraph
MLflow
W&B
Data & retrieval
PostgreSQL
Qdrant
Redis
Databricks
Spark
Airflow
Cloud & infrastructure
AWS
Google Cloud
Docker
Kubernetes
Terraform
GitHub Actions
Projects
Models we made smaller
We quantize open-weight models so they fit on less hardware, and publish every build on Hugging Face. Credit for each original model goes to the team that trained it.
Open-weight models on Hugging Face
3.5×
smaller than in bf16
~40 GB in 4-bit, from ~140 GB in bf16
Each square holds about 4 GB of weights.
LatamGPT 70B
The LatamGPT project's 70B model, built on Llama 3.1, with 4-bit weights and activations for Blackwell GPUs. We calibrated it on Spanish data on a single RTX 5090 workstation. Quality evaluations are still pending.
Based on LatamGPT (Llama 3.1 70B)Llama 3.1 Community License
2.9×
smaller than in bf16
8.25 GB in 4-bit, from 23.9 GB in bf16
Each square holds about 1 GB of weights.
humanizer
A model that rewrites AI-written drafts in English and Chinese so they read as human. We published two 4-bit builds: one that stays close to full precision, and one that trades fidelity for throughput on Blackwell's FP4 tensor cores.
hoja-solutions/
humanizer-NVFP4A16 4-bit weights, bf16 activations
94.77% next-token agreement with bf16
94.77%
next-token agreement with bf16
hoja-solutions/
humanizer-NVFP4 4-bit weights and activations
90.05% next-token agreement with bf16
90.05%
next-token agreement with bf16
Based on humanizer (Gemma 4 12B)Apache 2.0
Read our evaluationContact
Tell us about your project
Tell us what you are building. We will get back to you within one business day and be honest about whether we can help.