Play Deliberately

Agentic Architecture

This architecture integrates multiple specialized components designed to create a robust, intelligent system capable of sophisticated reasoning, execution, and continuous improvement. The architecture consists of four primary domains: Shared Memory, Planning, Action, and Reflection, Execution, and comprehensive Governance and MLOps frameworks.

Shared Memory

The foundation of our system’s intelligence is a sophisticated memory infrastructure that enables knowledge persistence and retrieval:

Data Lake: Forms the primary storage layer for raw, unprocessed data
Vector DB: Enables semantic search and similarity-based retrieval of information using vector embeddings
ML Features: Stores processed features ready for model consumption
Lake House: Provides structured data warehousing capabilities with the flexibility of data lakes

This shared memory system maintains bidirectional connections with the planning and action components, enabling improvement of the data as learnings happen.

Planning, Action, and Reflection

At the core of the system’s cognitive abilities is a sophisticated reasoning framework:

Chain of Thought: Implements step-by-step reasoning processes that break complex problems into manageable sequences
ReAct: Integrates reasoning and action in an iterative process, allowing the system to plan, act, observe, and adjust
Reflexion: Enables the system to evaluate its own outputs and performance, creating a feedback loop for continuous improvement
Tree of Thought: Extends chain-of-thought reasoning with branching decision paths, allowing exploration of multiple solution approaches simultaneously

Note that all of them need not be deployed at once. Initially Chain of Thought could be the one implemented and then adding the rest iteratively.

Execution

The execution layer transforms reasoning into concrete actions:

Agents: Autonomous LLM agents designed to perform specific tasks based on instructions from the planning layer
Classic ML Models: Traditional machine learning models that handle specific, well-defined tasks within the broader system

The execution layer interfaces directly with the Environment, which represents the external world or problem space where the system operates.

Governance

Our architecture incorporates robust governance controls to ensure responsible AI deployment:

AI Guardrails: Safety mechanisms that prevent harmful outputs or actions
Compliance: Features ensuring adherence to relevant regulations and standards
Utilization: Monitoring frameworks that track system usage and performance

MLOps

Supporting the entire architecture is a comprehensive MLOps framework that manages the operational aspects of machine learning deployment, including model versioning, deployment automation, and performance monitoring.

Other articles

Assigning Confidences to LLM Outputs

Published: Sat 18 January 2025
By Amit Agrawal

In Technical.

tags: Probability

LLM inferences tend to be erratically wrong. So, 99% of the time the answer is correct but 1% of the time it may be wrong and wrong in a way that is hard to predict and account for. In TruU we have technologies beyond just calibration for accounting for errors …
read more
Performance Capture

Published: Sun 05 January 2025
By Amit Agrawal

In Technical.

tags: Movies Photography

This image was my first foray into motion capture for Species in BOSS days. We all looked so young. SIL motion capture rig was all created internally at BOSS films. You can see Richard Edlund on the bottom right. This was really early days when we didn’t know what …
read more
Deep Names for Common Ideas

Published: Wed 23 October 2024
By Amit Agrawal

In Technical.

tags: Bayesian

Sometimes, we encounter complex names for concepts we’ve been using all along. Recently, I started exploring Dempster-Shafer Theory (more on that as I learn further) and came across the following:

Principle of Insufficient Reason

Also known as the Principle of Indifference, which, despite sounding like a bourgeois view of …
read more
Mandarin in Space

Published: Wed 23 May 2018
By Amit Agrawal

In Technical.

tags: VideoGames Education

Even though I have supervised developing of hundreds of apps, when boys were learning Mandarin in kindergarten, I decided to write this game myself to get them to practice their Mandarin. It is essentially a flash card game and actually can be extended to support any flash card implementations.

The …
read more
How does Gram Matrix encode the Style of an Image?

Published: Fri 01 December 2017
By Amit Agrawal

In Technical.

tags: Deep Network

In non-photoreal renderings, capturing the style of an image is an extremely tricky and difficult issue. Recently there has been a resurgence in this field with the application of Deep Learning to this problem. Specifically, in Gatys’ paper, the stylization problem is posed as an optimization problem where two cost …
read more
Nitty Gritty of Style Transfer

Published: Sat 25 November 2017
By Amit Agrawal

In Technical.

tags: Deep Networks

Style transfer was first introduced by Aaron Hertzmann’s classic paper on Image Analogies from SIGGRAPH 2001. More recently deep learning got applied to this problem in Leon Gatys paper “A Neural Algorithm of Artistic Style” which has led to a resurgence of work in this area. There are a …
read more
Playing with Gamma

Published: Tue 23 May 2017
By Amit Agrawal

In Technical.

tags: Movies Photography

Gamma is the non-linearity in the CRT which matches closely the non-linearity in the human eye. It is no wonder that all display devices now artificially embed the non- linearity in the display. For example, a DLP (digital light processing) display that is a linear display emulates the non-linearity of …
read more

Other articles

Principle of Insufficient Reason

links

social