AI/ML & DATA

AI/ML Development
and Data Science

Converging data and science turns theory into action through systematic data science & development processes.

How AI/ML and Data projects
work at CloudGeometry

Every AI/ML project is unique. In general, we find that customer engagements follow similar patterns, which lets us get you up and running in just a few weeks.
Our approach breaks the process down into stages to give you control over its progress to better manage its crucial feedback steps.

Discovery Workshop

Typically 8 hrs.
We start by spending a day with the business owners who are solving a problem, uncovering the details of the use case and what needs to happen to solve it successfully.

Results

Documented use case
Documented success metrics

Experimentation

Hosted Public
(e.g. OpenAI)
Self-hosted open source
(e.g. Llama)
Tuning existing models
Playing with technique
(e.g. RAG)

Building

Building apps
Creating & tuning models
MLOps for tracking experiments
Search company data for relevant info

Production

MLOps
for model drift
Retraining
if necessary
Observability
Managed hosting
OpenAI
Gemini
TensorFlow
PyTorch
Jupyter
Llama
HuggingFace
Kubeflow
MLflow — ML and GenAI
MLflow
Python
R Project
Argo
Grafana
Grafana
CloudGeometry
CloudGeometry works with your team throughout the three major phases of the machine learning process.

Pilot

Typically 4 weeks
The purpose of the pilot is to make sure that the project is feasible before you invest a lot of time, money, and resources. We start by defining what this will do for your business.  It’s only after you approve the Proof of Value that we proceed with the rest of the Pilot phase.
Once we've defined what this project will do and how it will help, we hold a 2 day kickoff workshop. This workshop is intended to not only educate your staff on any relevant concepts that might still be fuzzy, but to use our experience with existing models used by various industries to help you define both the model that will be best for you and the data to train it. If you don't have sufficient data, we'll define the data that needs to be generated. Also if you have unlabeled data, we'll define the process to label that data.
Finally, we'll create a minimal application that provides proof that the end goal is achievable.

Results

Proof of Value
Data & model definitions
Proof of concept
The CloudGeometry
AI/ML Blueprint
The machine learning process is a loop of continuous improvement.

Development: All stakeholders

Typically 4 weeks
Now we're ready to really get going and build your applications. In this stage, we'll be creating everything needed for your AI/ML systems to create the results you need. We'll be finding, creating, cleaning, and labeling data (if necessary), then training or tuning models as necessary. If it's a good fit, this is where we bring RAG best practices into the process.
This stage is where we help you create best practices for DataOps, MLOps, and/or LLMOps going forward. We'll also create everything needed for others inside and (if appropriate) outside your organization to use.

Results

Tuned/trained model, such as fine tuned LLM
DataOps / MLOps / LLMOps
API endpoint
Inference endpoint
Simple UI
Source code
User
User requests the solution to perform a task (e.g.: create a report)
Search
Search
Storage
Search company data for relevant info
Relevant
text
Send the most relevant info to LLM as context for performing tasks
Generate
Compose
LLM
Send instructions and info to LLM to get response
Task
output
Repeat the process as many times as necessary
We can help you use advanced techniques such as Retrieval Augmented Generation (RAG) to get better results from your applications.

Production

Typical transition: 3 weeks · available data maintenance & support
At this point, your application is built, and we help you prepare it for production, adding any necessary observability tools. We also help you transition to final hosting, such as cloud or managed hosting. Finally, we provide continuing support to ensure your models and applications continue to perform.

Results

Observability
Final hosting configuration

Continuing support

Watching for model drift
Retraining when necessary
Expanding application where desired