<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Aruna Chakkirala on Medium]]></title>
        <description><![CDATA[Stories by Aruna Chakkirala on Medium]]></description>
        <link>https://medium.com/@arunach321?source=rss-12f526612ac4------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*rqQ9cAIOyqSfNR62T93y6A.jpeg</url>
            <title>Stories by Aruna Chakkirala on Medium</title>
            <link>https://medium.com/@arunach321?source=rss-12f526612ac4------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 17:35:44 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@arunach321/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Skills instead of Tools for MCP]]></title>
            <link>https://medium.com/@arunach321/skills-instead-of-tools-for-mcp-fa1c268a1f3e?source=rss-12f526612ac4------2</link>
            <guid isPermaLink="false">https://medium.com/p/fa1c268a1f3e</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[mcp-server]]></category>
            <dc:creator><![CDATA[Aruna Chakkirala]]></dc:creator>
            <pubDate>Sun, 04 Jan 2026 16:52:36 GMT</pubDate>
            <atom:updated>2026-01-04T16:52:36.540Z</atom:updated>
            <content:encoded><![CDATA[<p>The recent paper from Anthropic has proposed using skills in lieu of tools for MCP. The primary purpose of MCP is to provide a universal protocol to connect Agents with tools and data sources. In the past year, multiple MCP servers have been build for different data sources, tools. Infact there are MCP registries like the official MCP Registry, github &amp; others for MCP server discovery.</p><p>An agent can have access to thousands of tools across multiple MCP servers. And therein lies the problem. These direct tool calls have mutiple challenges. For one, the tool definition constitutes the LLM context which translates to consumed tokens. When this number is extrapolated for the number of available tools, the tokens consumed increases exponentially. Secondly, the responses from each of the tool calls are shared as context in intermediate steps even when it is not needed while adding to the tokens consumed.</p><p>The recommended approach to handle the above challenges is to consider MCP servers code API’s. The server shares the list of tools it supports which could be represented in multiple ways. The paper suggests a file tree mechanism to expose the supported tools and the rest of the context is only loaded by the agent on need basis. So for instance if an agent requires the pdf reader tool, it only loads that tool details from the file (in the tree). The Agent then executes code (in a containerized enviornment) to interact with the tool and gets the results. The results can then be further processed to only extract data required for the agent.</p><p>The file tree structure mentioned above, fits in neatly into the skills functionality from a previous Anthropic paper (<a href="https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills">equipping-agents-for-the-real-world-with-agent-skills</a>).</p><p>The Skills.md file provides a metadata describing the supported skill/tool. The rest of the file only needs loaded when the tool is required. This reduces the context size to only tools which are required. The rest of Skills.md also allows for further nested skills thus allowing further decomposition and limiting the details loaded.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1021/1*-WcsddwUnqc5gXPSxDAPhA.png" /><figcaption>Source: <a href="https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills">https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills</a></figcaption></figure><p>Agents using code execution with MCP benefit from minimizing the tool definitions to need basis, preprocessing of tool results to only extract that which is required for the Agent and lastly privacy requirements (PII data) of the tool result data can be handled appropriately. State persistence is also a possibility as the tools tied to the filesystem can be used to store intermedate states. Additionaly just like in skills, over the course of an Agents lifetime the code genereated for a Skills.md file can be stored in the filesystem reducing regeneration time. Code is also preferred to handle simpler constructs like conditionals, loops when cost, latency is of concern.</p><p>Code execution with MCP addresses a number of the challenges from the previous tools passing method. Though there a few important factors to consider, the security of the code execution environment is key. Secondly access to the file systems should be controlled and also montiored for adverserial attacks. In addition, the Agent authentication, identification and validation will be an important element for security.</p><p>Original Papers:</p><p><a href="https://www.anthropic.com/engineering/code-execution-with-mcp">https://www.anthropic.com/engineering/code-execution-with-mcp</a></p><p><a href="https://blog.cloudflare.com/code-mode/">https://blog.cloudflare.com/code-mode/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fa1c268a1f3e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Deep Agents in Langchain]]></title>
            <link>https://medium.com/@arunach321/deep-agents-in-langchain-8b3b0d5e819e?source=rss-12f526612ac4------2</link>
            <guid isPermaLink="false">https://medium.com/p/8b3b0d5e819e</guid>
            <category><![CDATA[langchain-agents]]></category>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[ai-agents-in-action]]></category>
            <category><![CDATA[azureopenai]]></category>
            <category><![CDATA[aiagentsdevelopment]]></category>
            <dc:creator><![CDATA[Aruna Chakkirala]]></dc:creator>
            <pubDate>Sun, 30 Nov 2025 16:26:41 GMT</pubDate>
            <atom:updated>2025-11-30T16:26:41.713Z</atom:updated>
            <content:encoded><![CDATA[<p>Langchain v1 is an Agent first transformation of their popular framework. The framework is simplified with a basic agent creation as a starting point, the create_agent call provides a proven ReAct pattern on LangGraph’s durable runtime. The other addition is Deep agents for building complex, open ended long-term tasks.</p><p>The Agent-harness means that the framework comes preloaded with a specific set of core capibilities which are typical of complex Agentic Systems.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GPhXJniPXRJkKt4WZt1bcw.png" /><figcaption>source: <a href="https://docs.langchain.com/oss/python/deepagents/harness">https://docs.langchain.com</a></figcaption></figure><p>The built-in planner in Deep Agent is the “To-Do List” capability. The context management is handled with the “File System Tools” allowing agents to use persistent storage thus addressing context limitations arising from memory constraints. Specifically tools like web search are prone to provide large verbose variable length responses which can now The “Subagents” capability ensures isolated runs of subagents with their own individual contexts. Multiple subagents can function in parallel while enabling subagents to have different configurations and tools. There is also a “Long Term Memory” to store conversation context across threads (not to be confused with the File System Tools which is referring to intermediate context from tools or others).</p><p>The following code conducts research on a topic and the search queries are provided by the default planner of deep agent. Research is a classic example of benefitting from multiple parallel subagent runs to research on multiple sub topics before coming together to converge and collate. The model used here is Azure GPT5 and a filesystem backend has been added.</p><pre>tavily_client = TavilyClient(TAVILY_API_KEY)<br><br>def internet_search(query: str, <br>            topic: Literal[&quot;general&quot;, &quot;news&quot;, &quot;finance&quot;] = &quot;general&quot;,<br>            max_results: int = 5,<br>            include_raw_content:bool = False):<br>    &quot;&quot;&quot;Run a web search&quot;&quot;&quot;  <br>    response = tavily_client.search(query,max_results=max_results,<br>            include_raw_content=include_raw_content,<br>            topic=topic) <br>    return response<br><br><br>instructions = &quot;&quot;&quot;You are an expert researcher. Your job is to do a thorough research and publish a report. <br>A tool has been provided to you to search the internet and get the information<br>## &#39;internet_search&#39;<br>You can use this tool to run an internet search for a given query. You can specify<br>the maximum number of results, topic of the search and whether the raw <br>search content should be included in the results.&quot;&quot;&quot;<br><br><br>model = AzureChatOpenAI(model_name=AZURE_OPENAI_MODEL_NAME,  # or your deployment<br>    api_version=AZURE_OPENAI_API_VERSION,  # or your api version<br>    openai_api_key=AZURE_OPENAI_API_KEY,<br>    azure_endpoint=AZURE_OPENAI_ENDPOINT)<br><br><br>path = os.getenv(&quot;AGENT_ROOT_DIR&quot;)<br>deep_agent = create_deep_agent(backend=FilesystemBackend(root_dir=path),model=model,tools = [internet_search],<br>                    system_prompt=instructions)  <br><br>response = deep_agent.invoke({&quot;messages&quot;: [{&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: &quot;What are all the latest additions to Langchain&quot;}]})<br>print(response)<br><br></pre><p>Each of these capabilities also has a Middleware layering option to add any customization or controls. For instance the To-Do List can be updated with a Middleware which allows updating the To-Do List. This is a useful addition as previous Agent planners were difficult to be guided pointedly. Similarly the Sub-agent capability can be provided with user defined agents through the Middleware. In addition, the Sub Agent has access to a general-purpose agent which is primarily designed to keep the primary planner agent isolated from any request-responses and tool responses of the sub-agents.</p><p>Human in the loop insertion is better handled as an interrupt can be triggered for specific tool calls. An approval is then required by a Human reviewer for the tool execution. The Deep Agent has certainly simplified the Agent building framework while addressing specific pain points as well as organic learnings from diverse Agent building scenarios.</p><p>More details can be found on Langchain <a href="https://docs.langchain.com/oss/python/deepagents/overview">documentation</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8b3b0d5e819e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Evals driven AI Development — A Framework]]></title>
            <link>https://medium.com/@arunach321/evals-driven-ai-development-a-framework-c5729f4dba3d?source=rss-12f526612ac4------2</link>
            <guid isPermaLink="false">https://medium.com/p/c5729f4dba3d</guid>
            <dc:creator><![CDATA[Aruna Chakkirala]]></dc:creator>
            <pubDate>Tue, 16 Sep 2025 01:08:08 GMT</pubDate>
            <atom:updated>2025-09-16T04:03:33.514Z</atom:updated>
            <content:encoded><![CDATA[<h3>Evals driven AI Development</h3><p>Generative AI in enterprises has spurred a range of pilots across industries. Many organizations are aligned on the capabilities of these foundational models and the ensuing applications. Despite the consensus, only a small subset of the pilots make it into production.</p><p>A big reason cited is the availability of tools and platforms for comprehensive evaluations and observability. Evaluating GenAI applications is a key element in building trust and deploying into production.</p><p>Evaluation involves multiple factors, starting with ensuring that the model used is appropriate to the task on hand, the model’s accuracy right up to the validating the end-to-end application outputs.</p><p>Chip Hyu in her book, “AI Engineering” shares a framework for Evaluating applications. There is a lot of other literature also emphasizing the importance of robust evals. A step which is crucial in building trust and acceptance from the end users. And indeed ensuring the AI Applications can handle a range of expected queries with an acceptable accuracy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bN_u-3dQegI4ajOOWppCdw.png" /><figcaption>Azure AI Foundry Evaluations</figcaption></figure><p>There are metrics to monitor in eval phase. For instance in the Azure AI Foundry, the AI Quality is evaluated with metrics like Task Adherence, Intent resolution, Coherence, Fluency and Relevance. Each of these metrics targets a different aspect of the application. Relevance is checking the responses for being relevant to the query posed. While the metric ‘Task adherence’ is measuring on the responses being within an acceptable range of addressing the query without going astray.</p><p>These are generic evals which are useful but additional evals should also be added for each usecase. A subject matter expert is best equipped to augment the evals to address the specific task of the AI Agent. This could involve building a dataset of queries with expected responses which covers a wide range of the expected queries. In addition, Agentic workflows also require evals to track the Agents individually as well as the workflow of the mutliple agents working together. Each Agent should be evaluated on metrics discussed above alongwith others like appropriate tool calling, entity extraction etc. A planner agent within Multi Agent Systems should also be evaluated on its planning and orchestration abilities.</p><p>A common misconception is to rely on the model benchmarks as the final evaluation. Model benchmarks are indeed important for identifying the model for a use case but the success of the Agentic solution is reliant on the robustness and rigor in adopting an evals driven AI development.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c5729f4dba3d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Parallel Agents as a Patt]]></title>
            <link>https://medium.com/@arunach321/parallel-agents-as-a-patt-74b779efa019?source=rss-12f526612ac4------2</link>
            <guid isPermaLink="false">https://medium.com/p/74b779efa019</guid>
            <category><![CDATA[agent-design]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[genai]]></category>
            <dc:creator><![CDATA[Aruna Chakkirala]]></dc:creator>
            <pubDate>Mon, 01 Sep 2025 03:37:08 GMT</pubDate>
            <atom:updated>2025-09-16T01:09:33.799Z</atom:updated>
            <content:encoded><![CDATA[<h3>Parallel Agents</h3><p>In a previous blog we discussed emerging design patterns in building multi agent systems. Parallel Agents is gaining steam as a pattern for handling scale.</p><p>As the name implies, Agents works in parallel on the decomposed tasks while periodically consolidating or a supervisor agent assimilates the agents outputs. Concurrency is a key mechanism in managing scale; dsitributed systems or parallel programming being prime examples. The same analogy is now being extended to multi agent systems. Besides limitations on token length, time taken to generate next tokens or system latency are key factors for parallel agents.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*_uNUBrpVjnusEoP5e3Auqg.jpeg" /><figcaption>Parallel Agents</figcaption></figure><p>Deep Research is an instance of parallel agents, just like a human researches with mulitple tabs on a browser to gather multiple different elements before piecing it all together; the multiple agents work in a similar fashion. The research topic is divided across multiple agents and a master agent can collate it together.</p><p>Similarly coding can also be paralleized with agents where each agent is simulataneously building components a coding problem. The challenge of token limitations and the wait time in token generation is best handled here with parallelizing as many of the sub tasks.</p><p>Agent Design Patterns is constantly getting traction as better ways to solve real world problems are uncovered. The appropriate design of multi agent systems along with evals is a key element in the success of productionizing agentic solutions.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=74b779efa019" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Designing Multi Agent Systems]]></title>
            <link>https://medium.com/@arunach321/designing-multi-agent-systems-9300ab8ddfd9?source=rss-12f526612ac4------2</link>
            <guid isPermaLink="false">https://medium.com/p/9300ab8ddfd9</guid>
            <category><![CDATA[llm-agent]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[multi-agent-systems]]></category>
            <category><![CDATA[agentic-workflow]]></category>
            <dc:creator><![CDATA[Aruna Chakkirala]]></dc:creator>
            <pubDate>Sun, 23 Mar 2025 16:04:57 GMT</pubDate>
            <atom:updated>2025-03-24T02:48:04.235Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zF_EDSuuw2ABFFAnLts24A.jpeg" /><figcaption>Agents</figcaption></figure><p>“AI Agents will become the norm and humans will work alongside agents”, this is the rhetoric we are hearing from the industry.</p><p>So what is an AI Agent?</p><p>An Agent is the smallest unit of a collective where multiple agents can collaborate effectively to accomplish goals.</p><p>While a single agent can do any number of tasks, the ability of an Agent also resides in its armory or the tools it can use. A tool can be simply reading a document, writing a piece of code or even still running the piece of code. If tools are the arsenal of an Agent, the resources provide the knowledge or information source.</p><p>While a single agent can be useful, the real power lies in combining the power of many. Each agent takes up a task and collectively multiple agents collaborate to finish a goal. Interestingly enough, the collaboration can be cooperative or competitive. Agents can divide and conquer or they challenge the others leading to improvements.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1Z7MOee3J4IMfEH-_ONmqg.png" /><figcaption>Source: Multi-Agent Collaboration Mechanisms: A Survey of LLMs (2025) Tran et al</figcaption></figure><p>A recent survey paper on collaboration mechanisms shows the multiple Agent interaction methods. Cooperation, Competition and the third kind is Coopetition (a combination of cooperation &amp; competition).</p><p>Other factors include</p><ol><li>Environment/context provides the background for the goal.</li><li>The perception it can gather from the environment/context.</li><li>Actions which can be accomplished through tools.</li></ol><p>Agents can be backed by large language models or they can be rule based. Some Agent frameworks like autogen also use terms like “actors” to indicate a role based framework. The collaboration between the Agents is called workflows by many sources. There is another school of thought that workflows are more basic and hence different to Agentic interaction patterns. Despite the terminology variations, the Agent interaction design is an important detail in leading to a successful usecase.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9bLXUuBEOUxaRAH-ZHd3BQ.jpeg" /><figcaption>Agent Collaborations</figcaption></figure><p>Sequential is a series of Agents playing their part and then passing the baton forward. Gradually improving on the results, each Agent contributes to accomplishing the goal.</p><p>Hierarchical involves a orchestrator Agent assigning tasks to the appropriate agent. While a distributed interaction involves all agents working with or against each other. Human intervention is also an important addition ( human checkpoints can be added or one of the agents serves as a human proxy) when the interactions require human oversight. These are the basic design premises which have further evolved to tackle complex scenarios.</p><p>Most interactions will also include a planner agent to hatch out the plan for achieving a goal. Although the reasoning models are considered to render the role of a planner agent unnecessary. Additionally, security needs to be well thought as the tools have access to critical systems and their access has to be insync with the organizational governance practices. In this respect, integration through interfacing layers like MCP (Model Context Protocol) provide a layer of security and scalability.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*y4iOJb6hsPAJSx1h" /><figcaption>Photo by <a href="https://unsplash.com/@javaistan?utm_source=medium&amp;utm_medium=referral">Afif Ramdhasuma</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>A strong use case is undoubtedly the starting point for Multi Agent System. The second most important step is designing it right with the appropriate resources, tools and Agent interaction patterns. Thus enabling the multi agent systems to accomplish their <strong>goals</strong> effectively while maintaining security and accountability.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9300ab8ddfd9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The sparse advantage in MoE]]></title>
            <link>https://medium.com/@arunach321/the-sparse-advantage-in-moe-4affcc1d81e7?source=rss-12f526612ac4------2</link>
            <guid isPermaLink="false">https://medium.com/p/4affcc1d81e7</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Aruna Chakkirala]]></dc:creator>
            <pubDate>Mon, 12 Aug 2024 14:59:38 GMT</pubDate>
            <atom:updated>2024-08-20T04:10:16.521Z</atom:updated>
            <content:encoded><![CDATA[<p>Many models have joined the LLM ranks in this past year. Mixture of Experts (MoE) as a technique has been adopted in some models.</p><p>The first mention of mixture of experts was from a 1991 paper titled “Adaptive Mixture of Local Experts”. This can be considered akin to our brains where the appropriate region is used to recognize a face vs identifying places. In her book, “The worlds I see” — Fei Fei Li mentions an article by neuroscientist Simon Thorpe (The nature, 1996) which uncovers the speed of processing in human visual system. The underlying conclusion is that the specialized expert regions in our brain take over the image identification leading to exceptional speed. A similar stance is adopted by the Mixture of Experts (MoE).</p><p>Multiple models including Mixtral 8x7B fit this bill. Additionally Mixtral is termed as a Sparse Mixture of Experts (SMoE) language model. In Mixtral each layer has 8 experts. And for every token, a router network selects two experts. This is done at each layer, selecting the two appropriate experts, processing the current state and then combining their outputs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/733/1*Oz65YOjBU1X9ddN2zhn-qQ.png" /><figcaption>source: original paper from Mistral on MoE, <a href="https://arxiv.org/abs/2401.04088">https://arxiv.org/abs/2401.04088</a></figcaption></figure><p>The router in the architecture takes the responsibility of identifying the two experts.</p><p>The two selected experts leverage only a part of the model parameters unlike dense models where inferencing activates all the parameters. Each token only sees two experts, the selected experts can be different at each timestep. In the mixtral model, it has access to 47B parameters but it only uses 13B active parameters during inference. This leads to faster inferencing speed although a larger memory footprint is required to store all the parameters.</p><p>There are some interesting suggested alternatives for limiting the memory by storing the weights on a SSD &amp; only loading the weights required during the inferencing instance.</p><p>The training of the experts can also utilize efficient techniques like Sparse upcycling to address challenges of training large deep neural networks. In the paper titled, “Sparse Upcycling: Training Mixture-of-Experts from Dense Checkpoints” , the authors propose reusing training efforts of dense models by converting them into efficient sparsely activated Mixture-of-Experts (MoE) models. This method reduces the computational cost and data requirements in training the models. The paper also highlights that the upcycled models are more performant and efficient.</p><p>New efficient training techniques has garnered the interest of the research community and is an exciting space to watch.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4affcc1d81e7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Graph Analysis with Generative AI]]></title>
            <link>https://medium.com/@arunach321/graph-analysis-with-generative-ai-7be7a591dd2e?source=rss-12f526612ac4------2</link>
            <guid isPermaLink="false">https://medium.com/p/7be7a591dd2e</guid>
            <dc:creator><![CDATA[Aruna Chakkirala]]></dc:creator>
            <pubDate>Mon, 05 Feb 2024 15:05:52 GMT</pubDate>
            <atom:updated>2024-04-25T12:50:02.250Z</atom:updated>
            <content:encoded><![CDATA[<p>Graph Analysis has been prominent in data science for a few years now. A graph has vertices and edges where each vertex is an object and the edges denote the relationship between the vertices.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/650/1*gO1Oye3oPHoPc_-sIB_3kA.png" /><figcaption>source: <a href="https://www.cs.mtsu.edu/~xyang/3080/graph1.html">https://www.cs.mtsu.edu/~xyang/3080/graph1.html</a></figcaption></figure><p>In a business context, the graph depicts entities as vertices or nodes and their relationships is captured as edges. A graph is powerful when visualized since it captures multiple details in an instance and a human eye can quickly discern patterns from the visual. This is surely true with a small graph having few nodes and edges as shown in the figure below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/549/1*44rMEIWZEa7ywIAJ0sncsw.png" /><figcaption>Plotted Graph on customer names and their registered phone numbers</figcaption></figure><p>The graph above was created from customer details and their phone numbers. This information is typically stored in databases and finding patterns in data stored across many tables is tedious. In a enterprise context, graph analysis picked up steam in such scenarios.</p><p>Graphs have been around for many years in other areas like networks where it was predominantly used in routing algorithms. More recently it served the purpose of grouping similar objects in machine learning through clustering and identifying similar communities. For instance in social networks similar people are grouped and shared behavior&#39;s serve as a basis for the recommendations of articles to read or people to befriend.</p><p>The graphs used for identifying similar groups are unipartite since they have only one type of node; the nodes are user names in a social network. The figure above is a bipartite graph since there are two object types — names &amp; phone numbers. Bipartite or multipartite graphs are richer with additional detail on the relationships between the various node entities. A common use case for multipartite graphs is fraud detection. In the figure above, one can immediately discern that Alice &amp; Charlie are related in some sense as they share the same number. The example above is a toy example but when there are many people, the graph is richer. In most graph analysis use cases, a human will study the patterns and then build an automated manner to extract those patterns. This lengthens the time but quicker analysis is desirable for speedier action.</p><p>To handle this scenario, I was curious to check on Generative AI in identifying patterns from a graph. I did a quick experiment on Azure Machine Learning to load a graph (stored as an adjacency list, an efficient data structure for storing sparse graphs) and send it as a prompt to gpt 3.5 turbo.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5FmuX7mMLYvWD6nk4bUsVQ.png" /></figure><p>In the notebook, the Azure OpenAI instance (the Azure OpenAI service was created earlier from the Azure portal) was initialized and set to gpt3.5 turbo. Then a small adjacency list with names and phone numbers was used as context in the prompt, the query in the prompt was to determine all the relationships.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UU-_-zGQ7c6elqkJOLq54w.png" /></figure><p>The gpt model clearly explains the graph and rightly concludes that Alice &amp; Charlie, Charlie &amp; David are related .Interestingly, the LLM could also identify neighbors two hops away as show below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d4vErJ_wYpnpobC_u01avg.png" /></figure><p>This is a good example of the gpt’s reasoning power in determining a complex structure like a graph.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7be7a591dd2e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Causality with Large Language Models]]></title>
            <link>https://medium.com/@arunach321/causality-with-large-language-models-4829d2d8b2de?source=rss-12f526612ac4------2</link>
            <guid isPermaLink="false">https://medium.com/p/4829d2d8b2de</guid>
            <category><![CDATA[gpt-4]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[causal-inference]]></category>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[supply-chain]]></category>
            <dc:creator><![CDATA[Aruna Chakkirala]]></dc:creator>
            <pubDate>Mon, 15 Jan 2024 19:01:34 GMT</pubDate>
            <atom:updated>2024-01-17T16:52:56.514Z</atom:updated>
            <content:encoded><![CDATA[<p>Causal Inference is the process of identifying cause and effect between events. This has many applications across industries. Consider a supply chain optimization instance that involves identifying optimal routes for transporting goods. In such a scenario, the teams may have various questions on determining the optimal routes. What if I had used route B instead of route A? Will that cause a disruption in my supply chain? How can we reduce disruptions? This is best represented with a causal model for further interrogation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/256/1*nmLVc7Kuyz2ode3d39350w.png" /><figcaption>Created by Microsoft Designer. Prompt: Supply Chain Problems</figcaption></figure><p>The causal model fits into the second rung of the ladder of causation in Judea Pearl’s seminal work; ‘The book of why’. The ladder has three rungs, the lowest is association which can be determined from data with methods like correlation, predictions. The second rung refers to intervention which addresses the ‘What if I do’ or ‘How’ questions. The route in the above example is an intervention which we are performing to get answers for our ‘How’ questions. The third rung is the counterfactuals, posing retrospective questions — ‘What If I had done’ or ‘Why’. Counterfactuals can further be compounded with the presence of confounders like the weather, geo political factors which in turn can impact the ‘Why’ questions.</p><p>The entry of Large Language Models has drastically improved the accuracies for causal inferencing. The causal models which are typically depicted as a graph with nodes and edges provide a mechanism to answer ‘intervention’ questions with machine learning using techniques like Bayesian networks. Now with the entry of LLMs, models like GPT4 could answer the counterfactual questions. A <a href="https://arxiv.org/pdf/2305.00050.pdf">paper</a> titled “Causal Reasoning and Large Language Models: Opening a New Frontier for Causality” dated May 2023, has covered the LLM capability for Causal Reasoning. In one of trials, the model responded to a prompt requesting predictions on sales from data captured in the months of November &amp; December. The GPT4 model responds appropriately that the data is inadequate since December being holiday season could lead to increased sales. The GPT4 model then goes on to suggest that a further experiment be conducted.</p><p>The paper dives into details describing the prompting techniques which improved the models responses (a system message to the model calling out that it is a helpful assistant for causal reasoning)and also industry specific cause-effect relationships. Prompt Engineering also had an impact on the results as also the usage of appropriate phrases. The model also exhibited causal discovery capabilities with better accuracy and could also identify direct as well as indirect causes. I did a simulation of the scenarios listed in the paper on a GPT4 model with a prompt - to reason out the steps (called Chain of Thought, CoT method) and without CoT. Both responses correctly identified cause-effect relationships as show in the recording below.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FK_a8OHAi1KA%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DK_a8OHAi1KA&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FK_a8OHAi1KA%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/73625cc07a9c90f89f52eb6ff4c5d0a9/href">https://medium.com/media/73625cc07a9c90f89f52eb6ff4c5d0a9/href</a></iframe><p>It also rightly called out the absurdity in multiple counterfactual questions of the form — ‘what would happen if the planter grows in the plant’. When I tried this in the Azure AI Studio, GPT3.5 Turbo (with prompting to break into steps) made sense while GPT4 model gave the most plausible response.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qcfKD7PaDqJYFCeqW4d6pw.png" /><figcaption>Counterfactual example tried in Azure AI studio, GPT4 Response</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*U_OasE1L_JJbs8sWNOpSiA.png" /><figcaption>Counterfactual example tried in Azure AI studio, GPT3.5 Turbo with CoT</figcaption></figure><p>The model is essentially filling into the shoes of a human domain expert and also logically reasoning with techniques like Chain of Thought (CoT). The paper is from the early days of GPT3/GPT4, multiple other papers have since been published in this area. Investigating more examples will be essential to cover multiple scenarios, though the results from all the various papers look promising.</p><p>The causal reasoning ability of the large language models can open up an entire new area of integrations. A manufacturing unit which wants to troubleshoot to determine the cause of a fault or a sales team looking to right price the products. Across industries, the use cases for causality can benefit with LLMs in the midst as copilots.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4829d2d8b2de" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Gen AI as decoded by the experts]]></title>
            <link>https://medium.com/@arunach321/gen-ai-as-decoded-by-the-experts-1a931b99814f?source=rss-12f526612ac4------2</link>
            <guid isPermaLink="false">https://medium.com/p/1a931b99814f</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[genai]]></category>
            <dc:creator><![CDATA[Aruna Chakkirala]]></dc:creator>
            <pubDate>Thu, 28 Dec 2023 12:37:37 GMT</pubDate>
            <atom:updated>2023-12-28T17:01:53.409Z</atom:updated>
            <content:encoded><![CDATA[<p>With numerous Gen AI videos floating around, here are my recommended for the novice, curious &amp; the tech mind.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/312/1*1lK8DG-VeSiGezBFH8haDw.png" /><figcaption>Created by Microsoft Designer. <br>Prompt: Generative AI for the holidays</figcaption></figure><p>The Charlotte’s Web example below by <strong>Peter Lee (</strong>Microsoft Research Exec) is sharp &amp; succinct in quickly pointing out the power of the LLMs.</p><p><a href="https://www.linkedin.com/feed/update/urn:li:activity:7137158382624378880?updateEntityUrn=urn%3Ali%3Afs_updateV2%3A%28urn%3Ali%3Aactivity%3A7137158382624378880%2CFEED_DETAIL%2CEMPTY%2CDEFAULT%2Cfalse%29&amp;originTrackingId=Ls1yxF%2BLQAuEdAkCUd2fhg%3D%3D&amp;lipi=urn%3Ali%3Apage%3Ad_flagship3_profile_view_base_recent_activity_content_view%3Bcn03IfLnQHGMQKOireRtxA%3D%3D">Reid Hoffman on LinkedIn: A creative demonstration of the capabilities of GPT-4 through Charlotte&amp;#39;s... | 24 comments</a></p><p><strong>Micheal Wooldridge</strong> (Professor at University of Oxford) traces the significant milestones leading to Gen AI and masterfully explains the technicalities of large language models without getting technical.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fb76gsOSkHB4%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Db76gsOSkHB4&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fb76gsOSkHB4%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/f41996d9f0855d93d2282dd51a0c358e/href">https://medium.com/media/f41996d9f0855d93d2282dd51a0c358e/href</a></iframe><p>While <strong>Andrej Karpathy</strong> gets deeper into the workings of Large Language Models, scenarios and specifics around its usage.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FzjkBMFhNj_g%3Fstart%3D419%26feature%3Doembed%26start%3D419&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DzjkBMFhNj_g&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FzjkBMFhNj_g%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/4644c20f3265c0a23b9b98abb7c12e86/href">https://medium.com/media/4644c20f3265c0a23b9b98abb7c12e86/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1a931b99814f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Machine Learning lifecycle management]]></title>
            <link>https://medium.com/@arunach321/machine-learning-lifecycle-management-147500d62a0c?source=rss-12f526612ac4------2</link>
            <guid isPermaLink="false">https://medium.com/p/147500d62a0c</guid>
            <category><![CDATA[mlops]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Aruna Chakkirala]]></dc:creator>
            <pubDate>Sun, 17 Dec 2023 13:42:06 GMT</pubDate>
            <atom:updated>2024-02-05T15:25:36.774Z</atom:updated>
            <content:encoded><![CDATA[<p>An important aspect of building machine learning models is its iterative nature. Starting from the exploratory data analysis, model training until the model deployment - every phase benefits from multiple loops. The biggest challenge though in iterative processes is tracking the various runs. A good machine learning platform becomes imperative in such instances as it provides the ability to build models and also track the numerous runs.</p><p>The other important requirement for organizations centers around working collaboratively within a team; sharing models, components across environments while ensuring access is only provided on need basis. The goal of the good machine learning platform is to provide appropriate security controls while easing the process of building production ready models.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JK7nqIinUpuYk8CdWebmrA.png" /></figure><p>The snapshot above from Azure Machine Learning platform, displays the various models. Each job in the above visual translates to a model training (with specific dataset, model, parameters &amp; hyperparameters).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*u2U-kF5vdXpleSRBQMfl0A.png" /></figure><p>Digging deeper into a specific run, can lead to an end to end pipeline like above or be the details of a specific task. The snapshot below shows a training job along with the libraries, parameters, metrics observed and duration.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*c7Nxekn0U7dEHGx4t_MuRg.png" /></figure><p>As the data science team gets larger and more mature; it becomes an imperative to pick a Machine Learning platform. And the benefits of using a platform far outweigh the time and cost of setting it up.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=147500d62a0c" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>