#028 AI in Space: Solving Machine Learning’s Last Mile Challenge w/ Vid Jain

LISTEN: Podcast | Youtube | LinkedIn

LINKS → Vid Jain (LinkedIn) → Wallaroo.ai → “Startup Wallaroo Labs wins Space Force contract to model performance of AI on edge devices” (SpaceNews)

KEY POINTS

• Most companies wait too long to think about production. Start before you feel you need to do it.
• Issues to look for when operationalization AI models: 1) Production environment, 2) integration, and 3) data pipelines.
• ML Ops empowers data scientists to build better models, go faster, and limit business risks.
• Edge data/predictions must be incorporated into a central AI program to enable cross machine insights.

NOTES

• Data + AI = Value — not so fast. Not so easy because of deployment issues.
• Predictive analytics is a widely applicable use case. From space, to automotive and manufacturing.
• Wallaroo helps data scientists to operationalize AI. From the lab to a production environment.
• You can’t do “set and forget” with data science models. Requires constant monitoring.
• Issues to look for when operationalization AI models
  1. Production environment — what is the hardware that will be hosting the model.
  2. Integration — How will you model input data and output insights in a production setting?
  3. Data Pipelines — Do you have the data streams expected? How are you monitoring the health of those data pipelines?
• How MLops helps different personas
  1. Data scientists — spend more time on model development, rather than deployment/monitoring.
  2. Business Unit Leaders & MLOps engineers— Do more with less resources. Faster.
• Empower data scientists to
  1. Be bolder — build more interesting and sophisticated models.
  2. Go Faster — Shrink innovation cycle times.
  3. Limit Business Risk — Minimize translation errors in going from data science to production.
• Most companies wait too long to think about production. Start before you feel you need to do it.
• Appreciate that your core IP is in data and process. Monitoring and deployment, look for tools like Wallaroo.
• Thinking about benchmarking:
  1. “Bake-off” — Comparing models to one another. Swap between different models based on the value of predictions/insights.
  2. Raw Performance: Throughput, latency, etc. Compare across various production deployments.
• Edge vs. centralized managed AI. If you do inference at the edge, but aren't integrating back together in a centralized place in order to learn across deployments, you’re leaving “money on the table.”