# Notes
* https://www.qwak.com/post/mlops-pipeline-with-open-source-tools#5-machine-learning-workflow-orchestration
* https://neptune.ai/blog/mlops-tools-platforms-landscape
* https://towardsdatascience.com/scaling-pycaret-with-spark-or-dask-through-fugue-60bdc3ce133f
## Steps
1. Data readiness
2. Feature Engineering
3. Training and Hyper parameter tuning
4. Model serving
5. Model monitoring
1. Track model performance
2. Idenditfy Drift
3. Take corrective actions
7. Model management
## Functions
1. Schema at core
2. AutoML - Allow different framework to plugin
3. Feature Lake - Allow different implementation to plugin
4. Feature engineering pipeline
1. configurable
2. Allow Kafka vs Spark Pipeling (Batch vs Realtime)
3. Reproducability of features
4. Creation of new features
6. Training Pipeline - Spark based and other
7. Inference Pipeline - Batch and Real time
8. Data labeling pipline
## Implementation
1. Seldon Core - Model serving, AB Testing
2. Feast - feature lake
3. Evidently - evaluate, test, and monitor ML models
4. MLFLow - record experiment and store models
5. Kubeflow - ML @ scale
6. Fugue - scaling Panadas
7. Duck DB
8. Apache beam

## Components
### Feature Engineering Pipeline
1. Spark
2. Metadata
### Feature Store
https://redis.io/blog/feast-with-redis-tutorial-for-machine-learning/
https://github.com/feast-dev/feast/blob/master/protos/README.md
1. Feast
2. Parquet
3. Redis

### Knowledge

## About Whilter
https://fi.co/benchmarks

1. Current cash flow is managed via Racloop and this is working well for past 1 year.
2. ⁠We have 14 big logos with us currently. All of them are paying customers. So PMF is done.
3. ⁠This month order booking is 30-35 lakhs. We are targeting 70 lakhs revenue by six month.
4. A fund was ready to invest 4.5 Cr for 10% dilution 3 month ago when we had revenue of 15-20 lakhs. Since Racloop had enough cash flows we turn down this offer.
5. ⁠Its fund manager told us to reach **Golden MRR** of 70 lakhs(1 million ARR - leading) and we can get 20x valuation (20 million) for whilter. This is what we are chasing currently. They also suggested us to file patent and get ISO 27001.
6. ⁠We got ISO 127001 last month and and we are in process of filing patent.
7. ⁠We have current video generation cost at 40 paisa at 3 million scale and target is to get it down to below 20 paisa. According to meta this price will fly in US, Europe and Latin America.
8. ⁠Our GTM is very sorted - Meta, Gupshup, Yellow and Haptic are giving us customers as it is win win for them and us.
9. We are in process of doing POC with Brazil customer in portuguese via Gupshup.

## Working Togther
1. We are fine with acqution of Racloop by Whilter at minimal cost. We did a transaction at 6 Cr, 2 years back so I think we cannot go below it.
2. **We are betting big.** Whilter is the only priority for us, so we are not looking to make money from selling Racloop.
3. Our interest in AIONOS is to get capital infustion and scale faster so that we don't needed to dilute equity in future.
4. As mentioned yesterday we can treat two transaction seperately. AINOS invest in whilter seperately and we build Travel OS for AIONS seperately.
5. We have capabilities to build travel OS at 40% less cost and time. So working purely on BOOT(Build, Own, Operate and Transfer) model is a good business proposition.

## Immediate Changes
1. Internet Connection with backup
2. HR Practices - 2 people (Timesheet)
3. Payroll
4. Admin - 2 people
5. Infra + Laptops - Mac + Windows
6. Cab
7. GPU in house