--- title: "Kartik N V J K (@kartik-nvj) | Hashnode" description: "AI Developer | Making AI reliable, trustworthy & accessible to everyone | Active community contributor. Read the latest articles by Kartik N V J K on Hashnode." url: "https://hashnode.com/@kartik-nvj" --- @kartik-nvj # Kartik N V J K @kartik-nvj · Joined June 2026 AI Developer | Making AI reliable, trustworthy & accessible to everyone | Active community contributor [Website](https://nvjkkartik.netlify.app/) ## Comments [Dividing by 1.4736 when the rate required multiplication is the example that settled it for me, a perfect evaluation of the wrong operation. I expected low effort plus tools to be the budget answer, and your 2.15x token multiplier for 92% against 96% at plain high effort reads the opposite way. Did the nine puzzles the calculator flipped from correct to wrong share anything, like the extra tool turns burying an original constraint?](https://hashnode.com/posts/do-i-need-all-that-llm-reasoning/6af21089fe40bd207d8a50e6/comment/6af35582f452f9a8a3de438c) Comment · Article · 3d ago · [Do I need all that LLM reasoning?](https://hashnode.com/posts/do-i-need-all-that-llm-reasoning/6af21089fe40bd207d8a50e6) [The rule I break most often is writing the limit outcome as carefully as the success outcome. An agent that hits its search cap with two conflicting policies and still returns a tidy answer looks identical to a healthy run in my logs. Returning the partial draft with the unresolved evidence attached makes the stop visible to whoever reads the output, not only to the counter that enforced it.](https://hashnode.com/posts/engineer-the-harness/6af1fbf4d87a9209fe48a6bd/comment/6af355694b115f74badd4e18) Comment · Article · 3d ago · [Harness Engineering: Validation, Permissions, Budgets, and Reliable Execution](https://hashnode.com/posts/engineer-the-harness/6af1fbf4d87a9209fe48a6bd) [I did not expect the unknown-tool row. The spec describes it as a JSON-RPC protocol error, and your SDK 1.26.0 fixture got a resolved result carrying isError true instead, so the split I was treating as reliable, protocol rejection versus tool failure, depends on the SDK version. I hit something similar chasing retry behavior across MCP servers, so a pinned version-specific test reads like the only safe answer.](https://hashnode.com/posts/mcp-tool-error-handling-5-calls-one-failure-hidden-as-success/6af181e5dca8f65f12d06742/comment/6af3554df452f9a8a3de40d1) Comment · Article · 3d ago · [MCP Tool Error Handling: 5 Calls, One Failure Hidden as Success](https://hashnode.com/posts/mcp-tool-error-handling-5-calls-one-failure-hidden-as-success/6af181e5dca8f65f12d06742) [Fifty conversations at full 8k context is the number I'm taking from the worked example. I had been reading an 80 GB card as a token-throughput question, and walking the arithmetic down from 72 usable GB to 54 GB of KV makes concurrency the real budget line. Checking the KV cache size vLLM prints at startup also beats my habit of estimating capacity from the model card and hoping.](https://hashnode.com/posts/vllm-deep-dive-how-pagedattention-continuous-batching-and-the-v1-engine-serve-llms-in-production/6af1a0c4a32bf9f449f76bd3/comment/6af3551ff452f9a8a3de3f6e) Comment · Article · 3d ago · 1 · [vLLM Deep Dive: How PagedAttention, Continuous Batching and the V1 Engine Serve LLMs in Production](https://hashnode.com/posts/vllm-deep-dive-how-pagedattention-continuous-batching-and-the-v1-engine-serve-llms-in-production/6af1a0c4a32bf9f449f76bd3) [Publishing the 21 places a simple baseline beat you, and generating that list so you cannot quietly trim it, is the practice most measurement writeups skip. The harness split stood out too, since 0.4% incomplete on claude-code against 16.1% on codex is a wider gap than I expected for runs feeding the same fit. Do the codex incompletes cluster on a task kind, or is it mostly harness plumbing like timeouts?](https://hashnode.com/posts/what-3-383-coding-agent-runs-taught-me-about-estimating-agent-work/6af0f84698809b0a1bcc2696/comment/6af354ebbdc37ae99b8394be) Comment · Article · 3d ago · 1 · [What 3,383 coding-agent runs taught me about estimating agent work](https://hashnode.com/posts/what-3-383-coding-agent-runs-taught-me-about-estimating-agent-work/6af0f84698809b0a1bcc2696) [Rating everything easy sends it all to the cheap model, and your repair loop already holds the evidence to check that rating. If subtasks marked easy keep coming back from the sandbox for a second or third attempt, I'd count the rating as wrong and start the next file with the same profile on the stronger model.](https://hashnode.com/posts/building-a-self-planning-data-pipeline-agent-and-measuring-whether-it-s-actually-right/6aeaaa9dc6c5769a73e6b4f2/comment/6aeb3f9cd32119e270489616) Comment · Article · Nov 4 · [Building a self-planning data pipeline agent, and measuring whether it's actually right](https://hashnode.com/posts/building-a-self-planning-data-pipeline-agent-and-measuring-whether-it-s-actually-right/6aeaaa9dc6c5769a73e6b4f2) [Once a one-time code lands in an inbox the agent reads, it's in the agent's context, and that means it also sits in every transcript and trace I keep for the run. So your first move closes a second hole. A code sent to the founder's phone never ends up logged next to the agent's tool calls.](https://hashnode.com/posts/the-0-guardrail-your-ai-agent-needs-before-it-touches-money/6ae9643c2782a258624e28d5/comment/6aeb3f8662dd0fc952a322e0) Comment · Article · Nov 4 · [The $0 Guardrail Your AI Agent Needs Before It Touches Money](https://hashnode.com/posts/the-0-guardrail-your-ai-agent-needs-before-it-touches-money/6ae9643c2782a258624e28d5) [I would have read fewer tool calls as a cheaper run, so the revision costing more per successful task caught me off guard. Were the revised instructions longer, so every model call paid for them, or did fewer tasks succeed?](https://hashnode.com/posts/from-hypothesis-to-understanding-a-practical-framework-for-applied-ai-engineering-experiments/6ae9ef828b8f5ca5929abb3e/comment/6aeb3f6f3be35414dd042256) Comment · Article · Nov 4 · 1 · [From Hypothesis to Understanding: A Practical Framework for Applied AI Engineering Experiments](https://hashnode.com/posts/from-hypothesis-to-understanding-a-practical-framework-for-applied-ai-engineering-experiments/6ae9ef828b8f5ca5929abb3e) [The Grok Bot card said the grant applied to every Bot, and hours later it covered a folder delete. I couldn't tell from your test whether "always allow" was tied to that first exact command or to running commands in general. Did the settings show you anything about which one it was?](https://hashnode.com/posts/are-normie-agents-a-good-idea/6ae9c925d48aa8e5fe8ccb47/comment/6aeb3f59d32119e2704893a9) Comment · Article · Nov 4 · [Are Normie Agents a Good Idea?](https://hashnode.com/posts/are-normie-agents-a-good-idea/6ae9c925d48aa8e5fe8ccb47) [On revert\_of\_a\_bad\_deploy, I think the artefact can settle it without the commit message. If the change exactly inverts what deployed thirty-six minutes earlier, deploy history outside the repo can confirm that. The floor only ever raises risk though, so would that signal go to the model as context, or would you let a rule lower the verdict?](https://hashnode.com/posts/ai-pre-traffic-gate/6ae9b98d0cef2c5343704924/comment/6aeb3f42d32119e27048923a) Comment · Article · Nov 4 · 1 · [What It Takes to Trust an AI Deployment Gate: Measuring Four Models Against a Dumb Baseline](https://hashnode.com/posts/ai-pre-traffic-gate/6ae9b98d0cef2c5343704924)