Agentic AI

Creating an Agentic application is not the challenging part. There are many frameworks to do that, or you can rapidly make your own customized framework based on python classes. Do they RAG individually for having different personalities. This is required for truly constructive brainstorming within agents, do they orchastrate and iterate towards more optimized goals - thats where they act as thought amplifiers in an Al app. Then the LLMs (or SLMs) they go back to, ideally they should have the power of 'Reflection' and should be capable of advanced planning. The multiple round trips to LLMs have to be super fast.

One simply cannot use endpoints (APIs) of publicly hosted models. That would be too slow (and expensive) for anything other than toys apps. The infrastructure behind the Agentic framework is the real deal. Do you have your own LLMs optimized for speed, groundtruths, security, data safety, costing - various elements of LLMOps, in place. Well I will cover that in another micro post.

As an aside, in this age of Al, true value addition is in adding 'Innovation', which comes from a combination of deep domain knowledge and Al tech.

The source of such innovation can come from various things. My recent small pilgrimage to various places in India, culminating to our maha kumbh participation was one such booster - I can definitely think more holistically, and with equanimity and humility.