So I am building a new framework to later integrate into kruel.ai V7 Memory augmented network. This has me abit excited on the possibilities and will be another long trek I am sure.
Kainn: Kruel.ai Neurl Nets
Kainn is my experimental framework designed to push the boundaries of self-learning AI. By automating the processes of learning, testing, and refining, Kainn (kruel.ai Neural Nets) represents a significant step toward building AI systems that evolve autonomously and adapt to ever-changing needs.
What Kainn Does
-
Dynamic Learning: Kainn trains itself using provided data and dynamically integrates new information, allowing it to grow and improve without manual intervention.
-
Self-Testing: It evaluates its performance through continuous testing cycles, identifying gaps and opportunities for refinement.
-
Autonomous Improvement: Kainn uses feedback loops to adapt, retraining itself when errors are detected to ensure consistent performance enhancement.
-
Adaptability: Its learning capabilities enable it to respond to diverse and complex scenarios, evolving alongside the demands placed on it.
It’s in an early test phase right now to scope out the potentials of use cases. The first target is Intent systems to replace both static and NLP for something faster that consciously learns and adapts.
So far we have made a lot of progress and I am liking the results. once trained it can almost instantly produce intent or it will improve in that moment through self learning.
I have a feeling this will be almost as fun as it was to make the v7 memory which still is amazing and fastest memory concept I have built and tested. well Zzz time 5am comes early.
Update:
KAINN Intent Classification Framework - Automated Model Development Pipeline
A comprehensive framework for intent classification that automates the full ML lifecycle:
Features:
- Automated training data builder
- Autonomous test data generation and augmentation
- Self-optimizing training with edge case detection and corrections
- Automated validation targeting 99%+ accuracy
- Built-in deduplication and balanced sampling
- Model persistence and manual testing capabilities
This tool built on KAINN intelligently generates both positive and negative training examples, automatically identifies classification gaps, and performs targeted retraining until reaching specified performance thresholds. The underlying architecture ensures robust intent classification across diverse use cases while maintaining high accuracy standards.
Model speeds for response are in MS per transaction once fully trained.
This will be this months new testing tool with our memory data to check its performance across all its ai data points.
Auto Test Complete
Total Training Samples: 2023
Total Iterations: 10
Total Training Time: 0:05:50.726079
Accuracy Over Time:
Iteration 1: Accuracy: 100.00% - Misclassified Samples: 0 - Timestamp: 2024-12-03 08:31:40
Iteration 2: Accuracy: 100.00% - Misclassified Samples: 0 - Timestamp: 2024-12-03 08:31:58
Iteration 3: Accuracy: 97.32% - Misclassified Samples: 4 - Timestamp: 2024-12-03 08:32:30
Iteration 4: Accuracy: 96.62% - Misclassified Samples: 5 - Timestamp: 2024-12-03 08:33:08
Iteration 5: Accuracy: 98.62% - Misclassified Samples: 2 - Timestamp: 2024-12-03 08:33:43
Iteration 6: Accuracy: 98.11% - Misclassified Samples: 3 - Timestamp: 2024-12-03 08:34:27
Iteration 7: Accuracy: 97.96% - Misclassified Samples: 3 - Timestamp: 2024-12-03 08:35:07
Iteration 8: Accuracy: 100.00% - Misclassified Samples: 0 - Timestamp: 2024-12-03 08:35:46
Iteration 9: Accuracy: 98.70% - Misclassified Samples: 2 - Timestamp: 2024-12-03 08:36:21
We used a small sample just for demo purposes. This concept simplifies the build process. The back end uses Ai to build out the design to match what you are attempting to achieve with the simple classifications.
Generally my samples on a basic run start with 20k samples, You can also provide your own examples into the mix which the ai will look at to further build its edge / gap tests for further analysis and training to get the results with in spec.
