Surya Prakash Jayanti

TL;DR — I built a single AI agent that turns natural-language requests into well-structured tickets and generates weekly executive reports from the same source of truth. Result: cleaner data in, sharper reports out, less manual overhead. The Problem Engineering teams lose hours every week to two predictable workflows: Ticket intake — translating informal asks into properly linked, labeled, and scoped tickets. Weekly reporting — manually stitching status updates across goals, epics, and owners. The cost isn't just time. It's inconsistency — missing labels, orphan tickets, duplicate work, and reports that don't roll up cleanly. The Solution A single agent. Two skills. One source of truth. What Shipped CAPABILITY OUTCOME Unified agent One entry point — no more guessing which tool to use NL ticket creation Plain prompts → summary, description, epic link, labels, AC Duplicate detection Searches open tickets in the same epic before creating Label inheritance New stories auto-...

This week's AI pulse: The rapid ascent of agentic development is reshaping engineering workflows, bringing both powerful new capabilities and urgent calls for enhanced security and reliability to the forefront. 1. Know this AI Coding Agents Installing Unowned Packages Raise Security Concerns Why it matters: A recent report critically highlights that AI coding agents are inadvertently installing packages lacking clear ownership, exposing significant supply chain risks and creating critical accountability gaps within automated development pipelines. This development poses a direct threat to the integrity and security of production systems. Action: Immediately review and strengthen your organization's package management practices specifically for AI agents. Prioritize architecting AI platforms with robust, deterministic guardrails and rigorous evaluation pyramids to ensure scalability and reliability in production environments. 2. Try this Explore Datasette Agent for Co...

This week's AI pulse: With major model updates and new agentic tools arriving, the focus is sharpening on building reliable, secure, and highly capable AI-driven development workflows. 1. Know this AI Coding Agents Installing Untracked Packages Present Supply Chain Risks Why it matters: AI coding agents, in their quest for efficiency, are increasingly observed installing packages that lack clear ownership or oversight. This creates significant supply chain vulnerabilities, potentially introducing unvetted or malicious dependencies directly into your cloud-native and CI/CD pipelines without human accountability, as highlighted by Willem Delbare of Aikido Security. Action: Strengthen your CI/CD pipelines with automated package provenance checks and dependency scanning. Implement mandatory human review for all agent-generated dependency changes, ensuring every new component aligns with your organization's security policies and trust frameworks before deployment. 2. Try t...

This week's AI pulse: Agentic AI is rapidly maturing, moving from experimental scripts to governable, enterprise-ready solutions, backed by new models hitting general availability and integrated into core developer platforms. 1. Know this Google Rolls Out Gemini 3.5 Flash for Broad Developer and Enterprise Use Why it matters: Google's latest model, Gemini 3.5 Flash, has launched directly to general availability, skipping the typical preview modifier. Its immediate integration into billions of user-facing products like the Gemini app and Google Search, alongside developer platforms such as Antigravity, Google AI Studio, and Android Studio, signals a significant push for widespread AI model adoption across all segments, including enterprise via the Gemini Enterprise Agent Platform. Action: Investigate how powerful, broadly deployed models like Gemini 3.5 Flash are being baked into the developer tools and cloud-native platforms you rely on, and consider their implications ...

I really loved this book, I may go and find more books by author Mel Robbins. Here are some quotes that made me stop and read again. When you stop managing everyone else, you will realize you have a lot more power than you thought. You have just unknowingly been giving it away.  If your friends are not inviting you out to brunch this weekend, let them. If the person you have really attracted to is not interested in a commitment, let them. If your kids do not want to get up and go to that thing with you this week, let them. So much time and energy is wasted on forcing other people to match your expectations. The truth is, if somebody else you're dating, a business partner, a family member, is not showing up how you need them to show up, do not try to force them to change. Let them be themselves because they are revealing who they are to you. Just let them, and then you get to choose what you do next.  All human beings have a hardwired need for control.  Feeling in control ...

This week's AI pulse: The focus is squarely on bringing robust development practices and increased transparency to the rapidly evolving world of AI agents, with major platforms stepping up to support these sophisticated workflows. 1. Know this Cloudflare introduces Git-like versioning for AI agents Why it matters: Cloudflare's new Artifacts beta brings familiar Git-style version control to AI agent outputs, a critical step for managing, tracking, and evolving agentic systems with the same rigor as traditional codebases. Action: Start thinking about how you can integrate version control principles into your AI agent development and deployment pipelines to manage iteration and ensure reproducibility. 2. Try this OpenAI models expose interleaved reasoning in `llm` tool Why it matters: The `llm` 0.32a2 release for OpenAI models now allows developers to see summarized reasoning tokens when using GPT-5 class models via the `/v1/responses` endpoint, offering new debugging...

How I Built a Library-Constrained Book Recommender I read 200+ books. An agent picked my next 100 from my library shelf. I wanted recommendations constrained to books I can borrow now, not generic "you might also like" suggestions. So I built a pipeline that combines my reading history with Libby/OverDrive catalog data, then uses an LLM as a final curator. Code: github.com/jnsuryaprakash/book-recommender At a glance 210 prior reads parsed 16,450 library titles scanned 300 candidates to rerank 87 final picks Runtime: ~12 minutes local compute + ~30 seconds LLM call. LLM token cost: <$0.40 end-to-end. Output The pipeline produced themed, borrowable recommendations and stayed inexpensive to run (under $0.40 in LLM token cost end-to-end). Plano output: 87 picks across 9 themes. Sample themes includ...