About
I am a Principal Solutions Architect with 25 years of experience designing and delivering…
Articles by Sham Sundar
-
Building a Tiny AI in Your Browser (No Python Required)
Building a Tiny AI in Your Browser (No Python Required)
We are living through a massive platform shift. As AI transforms our industry, "upgrading our skills" has moved from a…
-
Why I paused the advanced AI rush to build a simple regression tool.
Why I paused the advanced AI rush to build a simple regression tool.
With the rapid pace of AI advancement, it feels like we are in a constant rush to learn the next big thing. But I’ve…
Activity
1K followers
-
Sham Sundar Hassan Chikkegowda shared this90 minutes on the clock. 4 teammates. One high-stakes internal innovation sprint. We placed 2nd. What got us there started before the clock did. When Palo Alto Networks brings cross-functional peers together under pressure, the dynamic changes. You're pitching to people who know the business, know the constraints, and have seen a hundred polished decks. That raises the bar. Before the sprint started, we focused on something most teams skip: communication frameworks. Rule of Three. Strip the jargon. Anchor every point in a business outcome, not a feature. That groundwork meant when the 90 minutes started, we weren't debating structure — we were building. Three things that actually mattered: 1). Speed through iteration. What separated the top teams wasn't who had the best idea at minute one — it was who refined fastest. Visual mockups, tested assumptions, story tightened right up to the last second. 2) Start with the human, not the solution. The moment our pitch landed hardest was when we stopped describing what we built and described the person who has the problem. One relatable observation. The room leaned in. 3). The Q&A is the real pitch. A polished 5-minute window is coachable. How you handle a sharp question from a VP under pressure is a different skill entirely — and it showed. Huge thanks to Rania Makriyiannis, Anjana Nair , Inzamam Sayyed and Tehreem Tungekar — the energy, visual creativity, and sharp thinking under pressure made this. And to the organizers — Alexandra Lee Mann, MM , Todd Hadley, Wendy S. and Edgardo Perez, Ed.M. — for designing a challenge that actually challenged us. The format, the time pressure, the executive panel — all of it was deliberate and it showed. Thank you to all, and the rest of the judging panel at Palo Alto Networks for the time, the sharp questions, and the feedback that's already shaping how I'll approach the next one. Walked in prepared. Competed hard. Placed 2nd — and I'm taking every framework from this sprint into the next room I pitch in.
-
Sham Sundar Hassan Chikkegowda shared thisDropped by the Fix My Agent clinic in SF at HogPatch in the evening non office hours, hosted by Courtney Ko and the teams at AI Valley, PostHog, GMI Cloud, MCPJam, Arcade.dev, and CodeRabbit. Even though I joined late (Though good welcome from Andrew Flores), here are a few key takeaways moving agents from fragile demos to production. • Sandbox Infrastructure (Mingjun Sun, GMI Cloud): Real production agents fall into three distinct lifecycles—sub-second ephemeral executions, long-running pipelines, and 24/7 digital employees. Long-running agents need checkpointing and state-resume capabilities so a crash 3 hours in doesn't blow your compute budget. • Protecting Website UI & Brand Rules : I asked Thierry Damiba how to prevent AI agents from breaking visual layouts when generating dynamic site content. The fix isn't prompt engineering—it's using gateway-level execution hooks and identity controls to restrict which tools or layout elements an agent can touch. • MCP Testing & Client Volatility (Prathmesh Patel, MCPJam): Standardizing tools with MCP is great, but LLM clients (like ChatGPT) frequently change endpoints or builds. Testing servers across multiple clients is becoming mandatory for reliability. • Job Hunting via MCP: Learned that Arcade built a dedicated MCP server where candidates' personal AI agents can apply directly to open engineering roles on their team. • Multimodal Maintenance Demo: Saw a practical hackathon pipeline where you upload an image of an electrical issue alongside a voice note, and a vision model returns step-by-step repair steps. Congrats Ali Amjad! • Code vs. Content: Interesting discussion on why software generation scales well with automated code reviews (like CodeRabbit), whereas "content factories" hit human attention limits where human creative direction remains the bottleneck. Good session focused on execution hooks, state recovery, and infrastructure boundaries rather than just bigger prompts.
-
Sham Sundar Hassan Chikkegowda shared thisA research team audited 75 AI-generated papers. 21% of the citations were completely invented. The writing looked convincing enough to pass. Shadab Hussain presented this at an AICamp session this week — covering Google's Co-Scientist, a multi-agent architecture built specifically to eliminate this problem. The session was technically precise, went well past the slides, and ended with a live debate on scientific integrity that nobody had planned for. One of the best hour-long sessions I've attended. I went home and built ClaimLens to apply the same idea to viral health hoaxes. Paste any headline. Three agents run: one decomposes the claim into specific testable parts, one hunts primary sources (.gov, PubMed, FDA records — not blog rewrites), one checks the evidence and delivers a verdict. A standard LLM runs on the same claim, same search results, in parallel — so the gap is visible on the same screen. Test: "Harvard study proves hot lemon water cures diabetes in 30 days." Standard approach: confident summary citing a Harvard study. 3 agents: no such study found. 9 of 10 results were health blogs echoing each other. Verdict: Fabricated. The metric I didn't plan for: the percentage of search results that are SEO noise with no primary source link — 87% on that claim. That number appears before the verdict and already tells you the search space is compromised. It showed up as a debug print and stayed in the app. Thank you Shadab Hussain for a session at that depth. And to AICamp for organising a talk that turned into a real debate nobody scheduled. Full writeup and live demo → links in comments.
-
Sham Sundar Hassan Chikkegowda shared thisA few hours ago I posted that I'd spent the first half of today volunteering, and the second half hacking on a problem I'd been sitting on. Here's what I built. I get a lot of messages on Slack and Discord. Real ones — from teammates, clients, collaborators — about topics I'm supposed to know about. Sometimes I do. Sometimes I don't. And the gap between "I got this message" and "I sent a good reply" is either an embarrassingly thin response or 20 minutes I didn't have. The root issue: there's no fast path from message received to I understand this well enough to reply properly. So I built Briefd — a personal reply agent that lives between your inbox and your reply. When a message arrives, it searches your personal knowledge base first. If it knows the topic, it drafts a reply with citations, pauses, and waits for you to approve before anything sends. If it doesn't know the topic, it spins up three research agents in parallel — web, Wikipedia, recent news — synthesizes what they find into a 5-point plain-English briefing, drafts from that fresh context, and again: nothing sends until you say so. You read the briefing. You read the draft. You click approve. It sends. Every reply you approve teaches it. Same topic next time — it already knows. What actually changed for me today wasn't the code. It was seeing it handle a topic it had never seen before, brief me on it in seconds, and produce a draft I would have sent myself. That's the moment it stops feeling like a tool and starts feeling like a colleague who did the reading so I didn't have to. Two things I didn't expect: The approval gate made me more willing to let the agent do more, not less. Knowing it cannot send without me reading the draft first is what makes it trustworthy enough to use on real messages. Deferred tool loading sounds like an implementation detail. It isn't. With four MCP servers connected, loading only the KB search tool at session start — everything else on demand — is what keeps a long session coherent. I wouldn't have known to care about this before today. The platform that made this possible is TrueFoundry. Their agent harness handled the parts that would have taken days to build from scratch: scheduling, subagent lifecycle, tool approval policies, Code Mode sandbox, context compaction across long sessions. I got real time with the people building it — Nikunj Bajaj, Sai krishna, Kashish Kumar Adam Chan. That access changed what I was able to ship in a day. If you get messages on topics you should know but don't always have context for — Slack, Discord, email threads — this is for you. The pattern: KB search → fast draft if known, briefing + research → draft if not → approval gate → send. It works for any domain. You bring the knowledge base; it handles the gap. More details and the writeup coming to rawweights.com. Built with TrueFoundry · OpenAI GPT-4o · Supabase pgvector · Exa · Next.js #TrueForge #TrueFoundry #Hackathon #BuildInPublic #AIBuilder
-
Sham Sundar Hassan Chikkegowda shared thisSpent the first few hours of today volunteering at the hackathon. Spent the second half as a hacker trying to solve my own problem. That shift hits different. When you're volunteering you're watching everyone else in the zone — heads down, something building. By midday I couldn't just watch anymore. I've been sitting on a problem that's been quietly annoying me for months. Today I finally had the time, the environment, and the energy to do something about it. Still early. Not ready to share what it is yet. But the feeling of going from "I help people build things" to "I'm building my thing, right now, in this room" — that's the one. Grateful to the team that put this together. The energy in a room full of builders is hard to replicate anywhere else. TrueFoundry HackerSquad
-
Sham Sundar Hassan Chikkegowda shared thisI asked an engineer whether their tool could serve as a legal compliance record. They said: "I don't know the right answer for that, actually." That answer — more than any demo — is what made the evening worth showing up for. I've been designing an agentic pipeline that runs under strict approval and compliance constraints. The hard problem: AI systems that run on their own can crash mid-task and duplicate their work, fall out of sync between what they write and what they can read back, or quietly touch more than they were supposed to — with nothing proving it either way. I spent an evening at a Builder After Hours session hosted by Temporal, MongoDB, and Keycard, stress-testing that design with the engineers who build this infrastructure for a living. The architecture has three layers, each owning exactly one problem: → Temporal — What happens when the server running your pipeline completely dies mid-task? This isn't a retry decorator. It's a different category of guarantee: the process can die and the work still finishes on a different machine, picking up exactly where it left off. → MongoDB Atlas + Voyage AI — Can the data your pipeline just wrote stay in sync with what your agent can search and read — without a separate sync layer that can fall behind or get out of step? → Keycard — Can you prove, after the fact, exactly what your agent accessed — and only what it was supposed to access? Two things I didn't expect: 1. The compliance record question. Event history in Temporal is an operational replay log. When I asked directly whether it could serve as a compliance artifact, the answer was an honest "I don't know." The redirect was to the right primitives: Signals to pause a workflow and wait for human approval, Queries to inspect what decisions were made. A meaningful distinction if you're building under regulatory requirements. 2. The trust boundary isn't the agent. I'd been thinking about access control as something applied directly to the agent. The Q&A corrected it: in an MCP architecture, the MCP server is the trust boundary — not the agent. The agent never holds credentials. That one correction changed how I'd design the entire access layer. Still in the design phase — sharing as I go. Full technical writeup with Python, MongoDB, and access-control code in the comments. Thanks for the direct answers: Melissa Herrera (Temporal) · Matte Noble (Keycard) · Nina Lopatina (MongoDB)
-
Sham Sundar Hassan Chikkegowda shared thisReading a good tutorial and actually starting the project are two different activities. The gap between them is where most ideas stop. Any AI copilot will summarise an article for you. But then you're still opening GitHub, creating a repo, pasting steps somewhere, and hoping the browser tab survives the week. I tried to wire up that whole chain — and have it live inside the workspace, not sit alongside it. Two ways to trigger it: open the web app and paste a URL, or type it directly into an Ambiguous channel. The agent picks it up, runs the pipeline, and posts the result back in the same thread. Either way, before anything gets created, you see the extracted steps and edit them. Same article, different idea typed in — completely different kit. That review step is deliberate. The human stays in the loop before the repo exists. After you confirm, one action creates everything inside Ambiguous — a channel message with the kit summary and GitHub link, a task, and a searchable document with every step. Not notifications sent to a separate tool. Artifacts that live where your work already is. GitHub gets the repo. Trigger.dev schedules a follow-up two days later: "did you start this?" Under 30 seconds. What broke along the way:Trigger.dev schedules a follow-up two days later: "did you start this?" Under 30 seconds. What broke along the way: CopilotKit powers the AI sidebar — ask it to list your channels, create a task, or search messages in plain English against your real workspace. I spent 90 minutes staring at a blank response. The function ran, the AI replied, nothing came back. No error anywhere. The library had quietly shifted to a new internal protocol and the old function silently swallowed the result. Four lines to fix once I found it. That's the part I always want to read in other people's writeups and rarely find. Full walkthrough in the blog. Full writeup, debug story, live demo in the comments Built with: Ambiguous · CopilotKit · Exa · OpenAI · OpenRouter · Trigger.dev · GitHub API
-
Sham Sundar Hassan Chikkegowda shared thisI bookmark 10 tutorials a week. I build maybe one a year. Not because I'm lazy. Because the gap between reading a great article and actually creating a repo, setting up the structure, and starting is just big enough to kill the momentum every single time. Yesterday I posted about exploring Ambiguous, CopilotKit, Exa and other sponsers — curious about what it could do as a workspace AI. Today that curiosity became a working product in a hackathon. I built BuildBridge: paste any technical article (or a YouTube video), and in under 30 seconds you get a GitHub repo, an implementation kit with editable steps, a Slack notification, a task in your workspace, and a follow-up reminder that nags you in 2 days if you haven't started. The whole thing runs on a human-in-the-loop pipeline — you see and edit every step the LLM extracted before anything gets committed. Because tools should augment your judgment, not bypass it. What surprised me building this: 1. Ambiguous is a coworker, not just a notifier. I wired it to fire three actions per kit — a channel message, a task, and a searchable document — all simultaneously. Then I added a CopilotKit sidebar so you can ask it questions in plain English. "List my channels." "Create a task for this kit with high priority." It turns a passive notification into an active workflow participant. 2. The optional idea field completely changes the output. Same FastAPI article, two different ideas → two completely different kits. One generic REST API tutorial. One with JWT tokens, bcrypt password hashing, and OAuth2 bearer auth. The LLM isn't just summarising — it's recontextualising the source for what you want to build. The part I didn't expect: pre-transcribing a 3-hour YouTube course with Whisper before the demo, and watching it produce 18 implementation steps covering everything from install to building a community manager overnight. The pipeline doesn't care if the source is an article or a video. It just works on the content. This is the kind of tool I would have used every week for the last three years. Big thanks to Jake Laes and Dasha Shunina — and the wider Ambiguous AI team. Managed a quick connection during the hackathon even with the internet doing its best to stop us. The Zoom didn't happen but the build did. Hoping to show you the working version properly. Built with #AgentsEverywhere AI Tinkerers, Ambiguous AI, OpenAI , CopilotKit , OpenRouter , Exa , Auth0 , Trigger.dev , Mozilla.ai , GitHub Next post: full video walkthrough + deeper look at how the pipeline works.
-
Sham Sundar Hassan Chikkegowda posted thisIn San Francisco, there's a room full of builders tommorow. Outside that room, the same hackathon is running in offices, kitchens, and co-working spaces across the world. I find that more interesting. It's Agents, Everywhere — a global hackathon organized by SF AI Tinkerers. No fixed idea yet. Just spent an evening on every sponsor's page. Here's what caught my attention from an evening of signup pages: OpenAI — Familiar ground. I've shipped with this. But every time I start a new project the question is the same: what's actually worth the API call, and what's just habit. CopilotKit — Used this in a previous build. The CoAgents piece — streaming live agent state directly into a React UI — is the part I haven't fully pushed. Curious what happens under time pressure. Exa — New to me, though I've used SerpAPI and QueryIt in pipelines before. Exa's framing is different: search designed for what a language model needs from results, not what humans click on. If that distinction holds up in practice, it changes how you think about research in an AI pipeline. Trigger.dev — "Background jobs that don't time out." I've hit that serverless wall at 10 seconds. This product exists because that frustration is everywhere. The question I want to answer: where are the actual edges? Ambiguous— 17 productivity apps rebuilt from scratch for human-AI collaboration. Not add-ons. Rebuilt. An agent with its own workspace identity raises more questions than the product page answers. That's usually a good sign. OpenRouter — One endpoint, every major model, route by task type. Makes intuitive sense. The interesting question isn't the API — it's whether thoughtful routing actually changes output quality in a way you can feel. Auth0 — I've used it for auth. Thinking about it as identity across multiple agent surfaces simultaneously — web, Slack, a workspace — is framing I haven't tried before. Mozilla — I use browser extensions daily. Never built one. MV3 apparently changed more than I expected. That gap between "I use this" and "I understand this" is exactly why I like these events. Not going in to win. Going in to experience what these products feel like under real constraints, alongside people who care about the same questions. If you're joining from your own office in a different city — not in the SF room — I'd genuinely like to know where you're building from. There's something worth paying attention to when the same problem runs in a dozen different contexts at the same time. To sponsor teams, fellow participants, and the SF AI community: looking forward to the exchange. The conversations usually matter more than the code. #AITinkerers #AgentsEverywhere #Hackathon #SanFrancisco #AI #BuildInPublic
Sham Sundar Hassan Chikkegowda commented on a post
21h
Congratulations Alexander Sivura! This is a great next step in your journey.
Our discussions around MD introduced me to you, agents, and eventually Data Science Farm. Your hustle and the way you approached agentic work motivated me, and I’ve learned a lot along the way.
Looking forward to seeing what you build and learn with agents in this new role at Google. No doubt you’ll rock it! 🚀
Sham Sundar Hassan Chikkegowda replied to a comment
3d
Car breakdown last weekend.
Stressed. Called insurance, arranged a tow, started updating everyone on affected plans. Standard protocol.
When I asked an AI assistant for help — it did the same. Followed the exact sequence I was already running.
Then I paused and asked it: is there any other option before the tow arrives?
There was a mechanic two minutes away. Car fixed before the tow truck showed up.
Same AI. Different question. The stress had narrowed my view — and the AI followed me into that narrow view.
Stepping outside the pressure, asking a different question — that was the human move.
Your point lands exactly: AI follows your framing. The irreducibly human skill is asking the question that breaks it.
Thank you Kevin T. for sharing the Neal Katyal story — thats the clearest example Ive seen of what AI supports, you decide actually looks like under pressure.
-
Sham Sundar Hassan Chikkegowda liked thisSham Sundar Hassan Chikkegowda liked thisI have always been a fan of open source. Maybe that started early in my career when I was at Adobe, working on the product that came from Day Software’s Communiqué (CQ) and eventually became Adobe Experience Manager (AEM). So when I got the chance to attend IndiaFOSS in Bengaluru, I was genuinely curious to see what people are building today. And there was a lot to see. 😄 I came across projects like: - NeST for network stack testing - Bruno, a Git-native API client - Frappe and its developer platform - Open infrastructure and payment projects - Projects around AI, security, hardware and much more But what I enjoyed most wasn’t any particular project. It was walking around and seeing people building things in the open. You can look at the code. You can try it. You can contribute. You can break it and figure out how it works. And maybe build something of your own on top of it. As someone who spends a lot of time around testing, developer tools, AI and agents, I find that incredibly interesting. I came back from IndiaFOSS with a simple thought: There is so much being built that most of us haven’t discovered yet. Now I have a few projects to explore. 🙂 If you were at IndiaFOSS, what caught your attention? And if you weren’t, tell me one open-source project you think I should check out. #IndiaFOSS #OpenSource #FOSS #SoftwareEngineering #DeveloperTools #AI #Testing
-
Sham Sundar Hassan Chikkegowda liked thisSham Sundar Hassan Chikkegowda liked thisThe people your organization needs most to lead AI transformation are probably not the ones currently inside it. That's not a dig. It's a structural reality. The people inside legacy organizations are spending every unit of cognitive and creative bandwidth just keeping up with the old world. Meetings. Deadlines. Reporting structures. A system designed to sustain itself, not reinvent itself. Meanwhile the people going deepest on AI right now? A lot of them are unemployed. Not because they failed. Because they have time. And they are using it to build genuine fluency in the new world without permission structures pulling them back. You cannot fully inhabit both worlds at once. Something gives. So here's my question for leaders. Are you creating real structural space for your people to cross over? Or are you extracting everything you can from the old world and calling it transformation? Because those are two very different things. #AITransformation #FutureOfWork #Leadership #ArtificialIntelligence #OrganizationalChange
-
Sham Sundar Hassan Chikkegowda liked thisSham Sundar Hassan Chikkegowda liked thisA little moment of gratitude from the past few weeks. ❤️ I was fortunate to receive two recognitions at Adobe for my work in the Agentic AI space. What made these especially meaningful was the feedback behind them. Being recognized for the work, while also hearing that my initiative, collaboration, and willingness to help have made a difference, meant a lot. There is so much to learn and explore in AI right now. Working in this space has given me the opportunity to take on new challenges, work through ambiguity, collaborate across teams, and keep learning along the way. It was really meaningful to receive this feedback and a good reminder to keep learning, contributing, and helping wherever I can. Grateful for the recognition, the trust, and all the people I get to learn and build with. 😊 #Adobe #AdobeLife #AdobeForAll #AgenticAI #SoftwareEngineering #ArtificialIntelligence
Experience & Education
Licenses & Certifications
-
M101J: MongoDB for Java Developers
Mongodb university
Issued -
CQ5.5 Lead Developer
Adobe
IssuedCredential ID 9A0-336 -
CQ5.5 Component Developer
Adobe
IssuedCredential ID 9A0-327
Volunteer Experience
-
urban agriculture initiatives
LEAF Urban Farm
- Present 9 months
Contributed to urban agriculture initiatives with LEAF Urban Farm, supporting sustainable farming and community food programs through regular Second Sunday volunteering.
-
Bike Repair
Bike Fremont
- 1 year 9 months
Volunteered with Fixing My Town in Newark/Fremont (2024–2025), assisting in Monthly community bike programs to promote mobility, sustainability, and local engagement.
-
Volunteer
Habitat for Humanity East Bay/Silicon Valley
- 1 year 9 months
Participated in home construction projects with Habitat for Humanity East Bay/Silicon Valley, contributing to affordable housing development through hands-on build activities.
-
Volunteer
internal community initiatives at Palo Alto Networks
Led and supported internal community initiatives at Palo Alto Networks, including back-to-school backpack drives and mentoring programs focused on interview preparation and career readiness.
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