AI Architect & Engineer | Agentic systems, RAG, AI infrastructure, Data Engineering | 738K+ LinkedIn, 294K+ Instagram | Newsletter for 250K AI builders
Charlotte Metro
740K followers
500+ connections
740K followers
500+ connections
About
I'm an AI Architect and hands on AI & Data engineer. I design the architecture, and I write the code needed to prove it, improve it, and ship it.
1M+ engineers, architects, and AI practitioners follow my work across Linkedin, X and Instagram. I publish practitioner level architecture breakdowns, visual explainers, and practical frameworks covering agentic AI, evals, MCP, AI infrastructure, and governance. 250K+ builders read my newsletter.
16+ years building secure, production grade AI and data platforms inside complex regulated enterprises across financial services, healthcare, and insurance. Patent holder in AI and fintech. Selected as one of 11 technical creators for the Oracle Creator Lab.
What I work on:
→ Agentic and multi agent systems: orchestration, planning, memory, durable execution and state, tool use, failure recovery, human in the loop, and AgentOps
→ Evaluation and observability: eval harnesses, LLM as judge, tracing, regression testing, drift and hallucination measurement, and closing the gap between benchmark performance and production behavior
→ Agent identity and security: verifiable identities for agents, users, tools, and workloads, delegated authorization, least privilege, consent, provenance, prompt injection defense, and zero trust controls
→ AI governance: policy enforcement, permissions, human approvals, red teaming, model and data jurisdiction, explainability, auditability, and Responsible AI
→ Enterprise RAG and GraphRAG: hybrid search, context engineering, query rewriting, retrieval optimization, grounding, and evaluation
→ AI infrastructure: multi model routing, inference cost and latency optimization, vLLM, BentoML, Kubernetes, GPU workloads, AI gateways, and hybrid local and cloud execution
Built with Python, FastAPI, LangGraph, Google ADK, Microsoft Agent Framework, MCP, vector databases, Kafka, Spark, Docker, Kubernetes, MLflow, AWS Bedrock, Vertex AI, and Azure AI.
I treat identity, governance, permissions, provenance, and trust as infrastructure. They cannot be documentation added after deployment. Every AI system should deliver measurable value, manageable risk, operational resilience, and sustainable total cost of ownership.
The next frontier is trusted infrastructure for intelligent workflows. That is where I'm focused.
Working with me: brand partnerships, sponsored technical content, developer marketing, and advisory for AI startups. Reach me at brij.py@gmail.com
Disclaimer: All views are my own and do not represent my employer.
Courses by Brij Kishore
-
Skill Sprint: Make the Most of Claude Code14m
Skill Sprint: Make the Most of Claude Code
By: Brij Pandey
Articles by Brij Kishore
Activity
740K followers
-
Brij Kishore Pandey reposted thisAn AI agent can give the right answer and still fail the test. Imagine a support agent that resolves a customer’s issue, but sends their personal data to an unauthorized tool along the way. If your evaluation only scores the response, that run could look successful. I’d want to know which identity made that call, what data the tool received, and whether we could have blocked the action or stopped the run. Those questions belong in the development process. That is part of what ADLC, the Agent Development Lifecycle, addresses. SDLC gives us established practices for building, testing, releasing, and maintaining software. MLOps adds versioned data, model evaluation, monitoring, and drift management. With agents, we also have to evaluate the tools and actions they choose during execution. Google Cloud’s Advent of Agents Season 3 opens with ADLC, then covers identity, security, deployment, observability, and operations throughout October. 𝗧𝗵𝗲 𝗳𝗶𝘃𝗲 𝗔𝗗𝗟𝗖 𝗽𝗵𝗮𝘀𝗲𝘀 -𝗦𝗰𝗼𝗽𝗲: Decide what the agent should do, who owns it, and where its authority ends. -𝗕𝘂𝗶𝗹𝗱: Develop with Agents CLI and ADK. Test its tool choices alongside its responses. -𝗦𝗰𝗮𝗹𝗲: Make deployment repeatable using infrastructure as code. -𝗚𝗼𝘃𝗲𝗿𝗻: Enforce identity, permissions, and policy throughout the lifecycle. -𝗢𝗽𝘁𝗶𝗺𝗶𝘇𝗲: Use evaluations and execution traces to understand failures and improve the next version. Governance starts when you define the boundaries. It does not wait until step four. 𝗕𝗲𝘆𝗼𝗻𝗱 𝗔𝗗𝗟𝗖, 𝘁𝗵𝗲 𝗺𝗼𝗻𝘁𝗵 𝗰𝗼𝘃𝗲𝗿𝘀 • Identity and access: how an agent acts on someone’s behalf without gaining extra permissions. • Runtime protection: handling prompt injection, data leakage, and code execution risks. • Fleet management: knowing who owns each agent and which tools it can use. • Observability and containment: reconstructing what happened during a run and stopping unsafe behavior. • Day 2 operations: managing cost, catching regressions, and continuing to evaluate after deployment. My suggestion: pick one agent and use it throughout the month. After each relevant episode, add a control, an evaluation, or a trace you can inspect. Each episode stands on its own, so start with a problem you are already working on. And keep the support-agent example in mind: would your current evaluations catch that unauthorized tool call? The series runs October 1 to 31, with one question a day, a practitioner video, and runnable code. It is open to everyone, with no sign in or paywall. Make this your October learning plan: 31 days of practitioner videos and runnable code on building, testing, securing, and operating AI agents. Pick a problem you’re working on and build along:https://fandf.co/4iK9wX9 Thanks to Google Cloud for sponsoring this post.
-
Brij Kishore Pandey shared thisAn AI agent can give the right answer and still fail the test. Imagine a support agent that resolves a customer’s issue, but sends their personal data to an unauthorized tool along the way. If your evaluation only scores the response, that run could look successful. I’d want to know which identity made that call, what data the tool received, and whether we could have blocked the action or stopped the run. Those questions belong in the development process. That is part of what ADLC, the Agent Development Lifecycle, addresses. SDLC gives us established practices for building, testing, releasing, and maintaining software. MLOps adds versioned data, model evaluation, monitoring, and drift management. With agents, we also have to evaluate the tools and actions they choose during execution. Google Cloud’s Advent of Agents Season 3 opens with ADLC, then covers identity, security, deployment, observability, and operations throughout October. 𝗧𝗵𝗲 𝗳𝗶𝘃𝗲 𝗔𝗗𝗟𝗖 𝗽𝗵𝗮𝘀𝗲𝘀 -𝗦𝗰𝗼𝗽𝗲: Decide what the agent should do, who owns it, and where its authority ends. -𝗕𝘂𝗶𝗹𝗱: Develop with Agents CLI and ADK. Test its tool choices alongside its responses. -𝗦𝗰𝗮𝗹𝗲: Make deployment repeatable using infrastructure as code. -𝗚𝗼𝘃𝗲𝗿𝗻: Enforce identity, permissions, and policy throughout the lifecycle. -𝗢𝗽𝘁𝗶𝗺𝗶𝘇𝗲: Use evaluations and execution traces to understand failures and improve the next version. Governance starts when you define the boundaries. It does not wait until step four. 𝗕𝗲𝘆𝗼𝗻𝗱 𝗔𝗗𝗟𝗖, 𝘁𝗵𝗲 𝗺𝗼𝗻𝘁𝗵 𝗰𝗼𝘃𝗲𝗿𝘀 • Identity and access: how an agent acts on someone’s behalf without gaining extra permissions. • Runtime protection: handling prompt injection, data leakage, and code execution risks. • Fleet management: knowing who owns each agent and which tools it can use. • Observability and containment: reconstructing what happened during a run and stopping unsafe behavior. • Day 2 operations: managing cost, catching regressions, and continuing to evaluate after deployment. My suggestion: pick one agent and use it throughout the month. After each relevant episode, add a control, an evaluation, or a trace you can inspect. Each episode stands on its own, so start with a problem you are already working on. And keep the support-agent example in mind: would your current evaluations catch that unauthorized tool call? The series runs October 1 to 31, with one question a day, a practitioner video, and runnable code. It is open to everyone, with no sign in or paywall. Make this your October learning plan: 31 days of practitioner videos and runnable code on building, testing, securing, and operating AI agents. Pick a problem you’re working on and build along:https://fandf.co/4iK9wX9 Thanks to Google Cloud for sponsoring this post.
-
Brij Kishore Pandey reposted thisYou type “What is gravity?” into an AI chat. ⠀ A moment later, an answer starts appearing: ⠀ “Gravity is a fundamental force that…” ⠀ Let’s follow that question through an LLM, from the text you type to the answer you read. ⠀ The model has already learned patterns during training. Using those learned patterns to generate an answer is called inference. ⠀ 𝟭. 𝗬𝗼𝘂𝗿 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻 𝗴𝗲𝘁𝘀 𝘀𝗽𝗹𝗶𝘁 𝗶𝗻𝘁𝗼 𝘁𝗼𝗸𝗲𝗻𝘀 ⠀ Tokens are small pieces of text. A token can be a whole word, part of a word or punctuation. ⠀ The carousel illustrates “gravity” as “grav” + “ity.” The actual split depends on the tokenizer. ⠀ Each token gets a numerical ID. ⠀ 𝟮. 𝗧𝗵𝗼𝘀𝗲 𝗜𝗗𝘀 𝗯𝗲𝗰𝗼𝗺𝗲 𝗹𝗶𝘀𝘁𝘀 𝗼𝗳 𝗻𝘂𝗺𝗯𝗲𝗿𝘀 ⠀ Each ID looks up a learned list of numbers called an embedding. ⠀ This gives the model a numerical representation it can work with. Information about token positions also helps it account for word order. ⠀ 𝟯. 𝗧𝗵𝗲 𝗺𝗼𝗱𝗲𝗹 𝗽𝗿𝗼𝗰𝗲𝘀𝘀𝗲𝘀 𝘁𝗵𝗲 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 ⠀ The numbers pass through transformer layers. ⠀ Inside each layer, attention lets tokens draw information from relevant tokens in the available context. Another component, the feed-forward network, further transforms each token’s representation. ⠀ These operations repeat across the layers, building representations that reflect the question. ⠀ 𝟰. 𝗜𝘁 𝘀𝗰𝗼𝗿𝗲𝘀 𝘄𝗵𝗮𝘁 𝗰𝗼𝘂𝗹𝗱 𝗰𝗼𝗺𝗲 𝗻𝗲𝘅𝘁 ⠀ The model’s output layer produces a score for each possible next token. Those scores are converted into probabilities. ⠀ For our question, the beginning of “Gravity” might be a likely starting point. ⠀ These probabilities describe possible text continuations. They don’t measure whether an answer is factually correct. ⠀ 𝟱. 𝗧𝗵𝗲 𝘀𝘆𝘀𝘁𝗲𝗺 𝘀𝗲𝗹𝗲𝗰𝘁𝘀 𝗮 𝘁𝗼𝗸𝗲𝗻 ⠀ It can choose the highest-probability token or sample from the possibilities. ⠀ Settings such as temperature influence that selection, which is one reason the same question can produce differently worded answers. ⠀ 𝟲. 𝗧𝗵𝗲 𝗽𝗿𝗼𝗰𝗲𝘀𝘀 𝗿𝗲𝗽𝗲𝗮𝘁𝘀 ⠀ The selected token becomes part of the context for predicting the next one. ⠀ First, the model processes the prompt. This is called prefill. ⠀ Then it continues generating. This is called decoding. ⠀ A KV cache saves intermediate information from earlier tokens so the model can reuse that work. ⠀ 𝟳. 𝗧𝗵𝗲 𝘁𝗼𝗸𝗲𝗻𝘀 𝗯𝗲𝗰𝗼𝗺𝗲 𝗿𝗲𝗮𝗱𝗮𝗯𝗹𝗲 𝘁𝗲𝘅𝘁 ⠀ Detokenization turns token IDs back into text. Streaming lets you see chunks of the answer while generation continues. ⠀ Some systems also use speculative decoding: a smaller model proposes several tokens, and the larger model checks them together. It’s an optional way to speed up generation. ⠀ The carousel follows this entire process using one question: “What is gravity?” ⠀ Once you can follow that example, terms like tokens, attention, sampling and KV cache become much easier to connect.
-
Brij Kishore Pandey shared thisYou type “What is gravity?” into an AI chat. ⠀ A moment later, an answer starts appearing: ⠀ “Gravity is a fundamental force that…” ⠀ Let’s follow that question through an LLM, from the text you type to the answer you read. ⠀ The model has already learned patterns during training. Using those learned patterns to generate an answer is called inference. ⠀ 𝟭. 𝗬𝗼𝘂𝗿 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻 𝗴𝗲𝘁𝘀 𝘀𝗽𝗹𝗶𝘁 𝗶𝗻𝘁𝗼 𝘁𝗼𝗸𝗲𝗻𝘀 ⠀ Tokens are small pieces of text. A token can be a whole word, part of a word or punctuation. ⠀ The carousel illustrates “gravity” as “grav” + “ity.” The actual split depends on the tokenizer. ⠀ Each token gets a numerical ID. ⠀ 𝟮. 𝗧𝗵𝗼𝘀𝗲 𝗜𝗗𝘀 𝗯𝗲𝗰𝗼𝗺𝗲 𝗹𝗶𝘀𝘁𝘀 𝗼𝗳 𝗻𝘂𝗺𝗯𝗲𝗿𝘀 ⠀ Each ID looks up a learned list of numbers called an embedding. ⠀ This gives the model a numerical representation it can work with. Information about token positions also helps it account for word order. ⠀ 𝟯. 𝗧𝗵𝗲 𝗺𝗼𝗱𝗲𝗹 𝗽𝗿𝗼𝗰𝗲𝘀𝘀𝗲𝘀 𝘁𝗵𝗲 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 ⠀ The numbers pass through transformer layers. ⠀ Inside each layer, attention lets tokens draw information from relevant tokens in the available context. Another component, the feed-forward network, further transforms each token’s representation. ⠀ These operations repeat across the layers, building representations that reflect the question. ⠀ 𝟰. 𝗜𝘁 𝘀𝗰𝗼𝗿𝗲𝘀 𝘄𝗵𝗮𝘁 𝗰𝗼𝘂𝗹𝗱 𝗰𝗼𝗺𝗲 𝗻𝗲𝘅𝘁 ⠀ The model’s output layer produces a score for each possible next token. Those scores are converted into probabilities. ⠀ For our question, the beginning of “Gravity” might be a likely starting point. ⠀ These probabilities describe possible text continuations. They don’t measure whether an answer is factually correct. ⠀ 𝟱. 𝗧𝗵𝗲 𝘀𝘆𝘀𝘁𝗲𝗺 𝘀𝗲𝗹𝗲𝗰𝘁𝘀 𝗮 𝘁𝗼𝗸𝗲𝗻 ⠀ It can choose the highest-probability token or sample from the possibilities. ⠀ Settings such as temperature influence that selection, which is one reason the same question can produce differently worded answers. ⠀ 𝟲. 𝗧𝗵𝗲 𝗽𝗿𝗼𝗰𝗲𝘀𝘀 𝗿𝗲𝗽𝗲𝗮𝘁𝘀 ⠀ The selected token becomes part of the context for predicting the next one. ⠀ First, the model processes the prompt. This is called prefill. ⠀ Then it continues generating. This is called decoding. ⠀ A KV cache saves intermediate information from earlier tokens so the model can reuse that work. ⠀ 𝟳. 𝗧𝗵𝗲 𝘁𝗼𝗸𝗲𝗻𝘀 𝗯𝗲𝗰𝗼𝗺𝗲 𝗿𝗲𝗮𝗱𝗮𝗯𝗹𝗲 𝘁𝗲𝘅𝘁 ⠀ Detokenization turns token IDs back into text. Streaming lets you see chunks of the answer while generation continues. ⠀ Some systems also use speculative decoding: a smaller model proposes several tokens, and the larger model checks them together. It’s an optional way to speed up generation. ⠀ The carousel follows this entire process using one question: “What is gravity?” ⠀ Once you can follow that example, terms like tokens, attention, sampling and KV cache become much easier to connect.
-
Brij Kishore Pandey reposted thisA lot of LLM calls inside agent pipelines exist to answer one small question. Which queue gets this ticket. How risky is this request. Should this tool call run. TypeSafe AI released Jev in early access on September 15, and it is built for those calls. It is a decision model and generates no text. The animation shows the timing. Jev returns three answers in a single beat. The LLM builds a sentence word by word, then hands it to a parser before the app can use it. 𝗛𝗼𝘄 𝗝𝗲𝘃 𝘄𝗼𝗿𝗸𝘀 You send state (a ticket, a record, some JSON) plus typed questions. There are three question types: - Choice picks one option from a set you define - Score places the input on an ordered scale and can land between levels, like 1.4 - Noul returns the probability that a yes or no question is true Every question is answered in one pass, each with probabilities. The answer can only be a value you defined, so there is no text to parse or repair. 𝗪𝗵𝗲𝗿𝗲 𝘁𝗵𝗲 𝗟𝗟𝗠 𝘀𝘁𝗶𝗹𝗹 𝗯𝗲𝗹𝗼𝗻𝗴𝘀 An LLM generates its reply token by token. You want that for reasoning, planning, writing, and code. For a routing decision it adds latency and cost, and your code still has to validate the text before acting on it. 𝗪𝗵𝘆 𝘁𝗵𝗶𝘀 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝗳𝗼𝗿 𝗴𝘂𝗮𝗿𝗱𝗿𝗮𝗶𝗹𝘀 Guardrails in an agent system are mostly decisions: should this action run, does this case need a human, is this output safe to send. A typed answer with a probability gives you a threshold you can tune, a value you can log for audit, and a clear rule for sending low confidence cases to a person. The policy itself stays in your code, where reviewers can read it. 𝗪𝗵𝗮𝘁 𝗜 𝘄𝗼𝘂𝗹𝗱 𝗰𝗵𝗲𝗰𝗸 𝗯𝗲𝗳𝗼𝗿𝗲 𝗮𝗱𝗼𝗽𝘁𝗶𝗻𝗴 𝗶𝘁 - TypeSafe has not published the model size, training data, or architecture - It is in early access, so pricing and limits may change - Vendor benchmarks tell you little about your own decisions. Pin a model version and test it against cases your team has already labeled If you were writing the rule for your team, what would decide whether a step in an agent gets a typed decision model or a full LLM?
-
Brij Kishore Pandey shared thisA lot of LLM calls inside agent pipelines exist to answer one small question. Which queue gets this ticket. How risky is this request. Should this tool call run. TypeSafe AI released Jev in early access on September 15, and it is built for those calls. It is a decision model and generates no text. The animation shows the timing. Jev returns three answers in a single beat. The LLM builds a sentence word by word, then hands it to a parser before the app can use it. 𝗛𝗼𝘄 𝗝𝗲𝘃 𝘄𝗼𝗿𝗸𝘀 You send state (a ticket, a record, some JSON) plus typed questions. There are three question types: - Choice picks one option from a set you define - Score places the input on an ordered scale and can land between levels, like 1.4 - Noul returns the probability that a yes or no question is true Every question is answered in one pass, each with probabilities. The answer can only be a value you defined, so there is no text to parse or repair. 𝗪𝗵𝗲𝗿𝗲 𝘁𝗵𝗲 𝗟𝗟𝗠 𝘀𝘁𝗶𝗹𝗹 𝗯𝗲𝗹𝗼𝗻𝗴𝘀 An LLM generates its reply token by token. You want that for reasoning, planning, writing, and code. For a routing decision it adds latency and cost, and your code still has to validate the text before acting on it. 𝗪𝗵𝘆 𝘁𝗵𝗶𝘀 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝗳𝗼𝗿 𝗴𝘂𝗮𝗿𝗱𝗿𝗮𝗶𝗹𝘀 Guardrails in an agent system are mostly decisions: should this action run, does this case need a human, is this output safe to send. A typed answer with a probability gives you a threshold you can tune, a value you can log for audit, and a clear rule for sending low confidence cases to a person. The policy itself stays in your code, where reviewers can read it. 𝗪𝗵𝗮𝘁 𝗜 𝘄𝗼𝘂𝗹𝗱 𝗰𝗵𝗲𝗰𝗸 𝗯𝗲𝗳𝗼𝗿𝗲 𝗮𝗱𝗼𝗽𝘁𝗶𝗻𝗴 𝗶𝘁 - TypeSafe has not published the model size, training data, or architecture - It is in early access, so pricing and limits may change - Vendor benchmarks tell you little about your own decisions. Pin a model version and test it against cases your team has already labeled If you were writing the rule for your team, what would decide whether a step in an agent gets a typed decision model or a full LLM?
-
Brij Kishore Pandey reposted thisBrij Kishore Pandey reposted thisIf you want to actually understand LLMs and Generative AI (not just copy prompts), this Stanford lecture series is one of the cleanest step-by-step paths I’ve found. Here are the 9 sessions (watch in order): Transformer Fundamentals — https://lnkd.in/gaaRDexT Transformer Models + Practical Tricks — https://lnkd.in/gy4FUwNY Transformers → Large Language Models — https://lnkd.in/gsPiCrEU How LLMs Are Trained — https://lnkd.in/gvHJvgqP Tuning & Adaptation (fine-tuning, etc.) — https://lnkd.in/g6kgtPKR Reasoning in LLMs — https://lnkd.in/gAACSUG6 Agentic LLMs (tools, planning, workflows) — https://lnkd.in/gVm6js9z Evaluation: what “good” really means — https://lnkd.in/gJhbFQ4s Recap + What’s trending now — https://lnkd.in/g5JMNTsf Also, subscribe to this - https://appliedstack.ai/ Learn AI by doing - and https://brij.media/zero My suggestion: treat it like a mini-bootcamp, one lecture at a time, take notes, and pause to implement small experiments after each video. If you’re learning LLMs in 2026, this is a solid foundation.
-
Brij Kishore Pandey reposted thisThis diagram shows every part of an AI agent at work. This animation follows one AI agent through a full run. The request is "Summarize Q3 churn, email the team." The agent goes around the plan, act, observe loop three times before the task is done. Each dot color shows what is moving: - Orange: the task and every command - Blue: results coming back from a tool - Pink: short term memory - Yellow: long term memory - Lime: trace events 𝗪𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝘀 𝗶𝗻 𝗼𝗻𝗲 𝗿𝘂𝗻 Lap 1 runs a SQL query for the churn numbers. Lap 2 runs Python to build the chart. Lap 3 sends the email, and that step waits at the guardrails until a person approves it. At every PLAN step the LLM reads short term memory to see what it has already done. At every OBSERVE step the result goes to long term memory, and a trace event lands in the observability panel. 𝗧𝗵𝗲 𝗽𝗮𝗿𝘁𝘀 - Instructions set the goal and the limits before the first step - The LLM chooses the next step and hands the action to a tool - Memory carries context between steps and between runs - Tools act on real systems: a database, a code runtime, an inbox - Guardrails check every tool call before it runs - Traces record each step so someone can review the run later 𝗪𝗵𝗲𝗿𝗲 𝗜 𝗽𝘂𝘁 𝘁𝗵𝗲 𝗮𝗽𝗽𝗿𝗼𝘃𝗮𝗹 𝗹𝗶𝗻𝗲 The SQL query and the Python chart run on their own. They read data and produce a draft. The email reaches other people and cannot be pulled back once sent, so it waits for a person. I sort agent actions into three levels by what happens if they go wrong: reading data, changing internal records, and anything that leaves the organization. Each level gets a stricter check. Keeping that rule in the guardrail layer puts the policy in one place, where reviewers can read it, instead of spreading it across prompts. 𝗪𝗵𝘆 𝘁𝗿𝗮𝗰𝗲𝘀 𝗺𝗮𝘁𝘁𝗲𝗿 When an agent sends the wrong report, the first question is what it saw and why it picked that step. A trace that records the plan, each tool call, each result, and each approval answers that question. With only the final output, the team is left reconstructing the run by hand. Which agent actions do you think should always need a human sign off, however capable the models become?
-
Brij Kishore Pandey shared thisThis diagram shows every part of an AI agent at work. This animation follows one AI agent through a full run. The request is "Summarize Q3 churn, email the team." The agent goes around the plan, act, observe loop three times before the task is done. Each dot color shows what is moving: - Orange: the task and every command - Blue: results coming back from a tool - Pink: short term memory - Yellow: long term memory - Lime: trace events 𝗪𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝘀 𝗶𝗻 𝗼𝗻𝗲 𝗿𝘂𝗻 Lap 1 runs a SQL query for the churn numbers. Lap 2 runs Python to build the chart. Lap 3 sends the email, and that step waits at the guardrails until a person approves it. At every PLAN step the LLM reads short term memory to see what it has already done. At every OBSERVE step the result goes to long term memory, and a trace event lands in the observability panel. 𝗧𝗵𝗲 𝗽𝗮𝗿𝘁𝘀 - Instructions set the goal and the limits before the first step - The LLM chooses the next step and hands the action to a tool - Memory carries context between steps and between runs - Tools act on real systems: a database, a code runtime, an inbox - Guardrails check every tool call before it runs - Traces record each step so someone can review the run later 𝗪𝗵𝗲𝗿𝗲 𝗜 𝗽𝘂𝘁 𝘁𝗵𝗲 𝗮𝗽𝗽𝗿𝗼𝘃𝗮𝗹 𝗹𝗶𝗻𝗲 The SQL query and the Python chart run on their own. They read data and produce a draft. The email reaches other people and cannot be pulled back once sent, so it waits for a person. I sort agent actions into three levels by what happens if they go wrong: reading data, changing internal records, and anything that leaves the organization. Each level gets a stricter check. Keeping that rule in the guardrail layer puts the policy in one place, where reviewers can read it, instead of spreading it across prompts. 𝗪𝗵𝘆 𝘁𝗿𝗮𝗰𝗲𝘀 𝗺𝗮𝘁𝘁𝗲𝗿 When an agent sends the wrong report, the first question is what it saw and why it picked that step. A trace that records the plan, each tool call, each result, and each approval answers that question. With only the final output, the team is left reconstructing the run by hand. Which agent actions do you think should always need a human sign off, however capable the models become?
Brij Kishore Pandey commented on a post
2h
Congratulations Nandini Ramani!
-
Brij Kishore Pandey liked thisBrij Kishore Pandey liked thisToday I am joining New Relic as Chief Executive Officer and member of the Board of Directors. I have spent my career working on the infrastructure problems that sit closest to how engineers build and operate software. As AI transforms how applications are built and operated, engineers need observability tools they can trust, and New Relic is built to deliver that. New Relic sits at an extraordinary inflection point in enterprise technology. As AI transforms application architectures and complexity escalates, engineers need observability tools they can trust implicitly, and New Relic is built to deliver exactly that. I have dedicated my career to solving this problem, and I am thrilled to lead this talented team as we expand our AI-powered observability platform and help our customers build with clarity and confidence. Thank you to the New Relic team for the warm welcome. I look forward to working with each of you and am thrilled to lead such a talented team.
-
Brij Kishore Pandey liked thisBrij Kishore Pandey liked thisAI advances faster when diverse minds are part of the conversation. On International Women in AI Day, LuMay AI proudly celebrates the women who bring curiosity, empathy, creativity, technical excellence, and relentless execution to everything we build. At LuMay AI, an inclusive culture is more than a principle. It means creating equal opportunities to learn, lead, experiment, contribute, and make a meaningful impact. From shaping AI products and solving customer challenges to strengthening our teams and culture, our women colleagues play an important role in the LuMay journey every day. To the incredible women of LuMay AI: thank you for your passion, dedication, ideas, and the impact you continue to create. Here’s to building an AI future where more women have the opportunity to lead, innovate, and shape what comes next. Happy International Women in AI Day! 💜 Surekha Chikoti Mary Grygleski Reshma Sriraman Bioshini Elizabeth V Anu Priya R Dhanusha Sivasankari Ponnusamy #WomenInAI #InternationalWomenInAIDay #WomenInTech #ArtificialIntelligence #AILeadership #WomenInTechnology #InclusiveInnovation #LuMayAI Women AI Builders
-
Brij Kishore Pandey liked thisBrij Kishore Pandey liked thisEnjoyed a wonderful trip to Norway 🇳🇴 with my 81 year old mother. Marleen van Heddeghem - Schepens ❤️
-
Brij Kishore Pandey liked thisBrij Kishore Pandey liked thisIf you missed LinkedIn's NewFront last week - here’s 40 seconds on buyability! In B2B, reach and other vanity metrics alone don’t move businesses forward. What drives growth is whether your business is remembered, trusted and chosen. That's buyability - the confidence a buying group feels leading up to a purchase. 👉 Brands are 20 times more likely to be purchased when the whole buyer group has awareness of the brand 👉 Buyers are also four times more likely to buy from a vendor they've had previous success with Check out the LinkedIn Virtual Newfront to learn more about how to become buyable using full-funnel video solutions: https://lnkd.in/eKTkV5KD
-
Brij Kishore Pandey liked thisBrij Kishore Pandey liked thisA few months after joining LinkedIn, I was struggling to adjust from entrepreneur to pm at a larger company. I went to Ryan’s office (a fast-rising exec at the time with this killer profile) and let him know it was best for everyone if I moved on. The next morning, I came into work and there was Ryan, sitting right next to me in our little shared cubicle. He told me he had decided to change desks. I learned more watching him in those next few months, and had more fun, than I ever have at work. And thanks to his act of kindness, I ended up not quitting. Ryan Roslansky, you’re a one of a kind leader and friend. Thank you for always setting the bar on doing well & doing good for each other.
-
Brij Kishore Pandey liked thisBrij Kishore Pandey liked thisA friend asked me: “We gave everyone coding agents. Why are changes taking longer to ship?” I couldnt answer on the spot, but it turned out to be classic queueing theory. Agents increased the number of PRs arriving (producers), but review capacity stayed flat (consumers). When a team is already near full utilization, even a small increase in arrivals creates a much longer wait. Once PRs arrive faster than people can review them, the backlog never clears. PRs opened per day gt; PRs reviewed per day → backlog grows every day On top of that, one agent task can also produce several related PRs across different repositories. Handle those PRs independently, and the team reconstructs the same context for every piece. Then you hear that manager telling you to review faster, which, spoiler alert, isnt much of a strategy. I think we need to focus on giving that time back by making priorities clear and keeping related changes together. That’s the problem Qodo 3.0’s PR Triage goes after. It groups related PRs into a single work package, ranks it by impact and open findings, and shows every repo it touches. You still review the code, but you see the whole unit of work instead of rebuilding it PR by PR. Faster code generation without faster verification just creates a longer queue. https://lnkd.in/e53AVVsK
-
Brij Kishore Pandey liked thisBrij Kishore Pandey liked thisThe Roman Empire paid for Indian goods in gold, and a single hoard in southern India held over 8,000 of the coins. The Kottayam hoard, north of Nelcynda, was probably deposited from a Roman vessel that foundered off the coast, and the latest coins were struck under Nero. More than 20 hoards from the reign of Augustus alone have turned up in the Tamil country, the Coimbatore Gap has produced more Roman coins than the rest of the subcontinent combined, and the Chera, Pandya and Chola kings minted coins imitating Roman portraits. Pliny put the annual outflow at 50 million sesterces. Kottayam, at about 800,000, is a small fraction of a single year. I've written up the full story in my latest Substack essay (link below).
-
Brij Kishore Pandey liked thisBrij Kishore Pandey liked thisExcited to share that I’ve been selected as an #IBMTechVoice A selective technical advocacy program bringing together IBM’s technical creators and voices who share knowledge, insights, and emerging technology with the developer community. I am pretty much grateful to be an IBM Tech Voice alongside my work as a Research Engineer at IBM Research. What makes this especially meaningful to me is that it complements something I already love doing: Learn → Build → Explore → Share Through my research work, I get to work on interesting AI problems and explore how emerging technologies actually work. Through TechVoice, I get another opportunity to share those learnings, simplify complex technical concepts, connect with the developer community, and learn from other technical voices across IBM. Looking forward to learning more, creating more, and contributing to the conversations shaping AI and technology. Grateful for the opportunity. Excited for what’s ahead. 💙 ------------------------- Hi, I am Aakriti Aggarwal, Research Engineer at IBM Research. Making complex topics → Easy.
Experience
Education
Languages
-
Hindi
Native or bilingual proficiency
-
Bhojpuri
Native or bilingual proficiency
-
German
Elementary proficiency
-
English
Native or bilingual proficiency
Other similar profiles
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content