A single determined builder can now go from idea to working product in days, sometimes hours. The bottleneck is no longer access to technology. It's clarity.
The bottleneck is no longer access to technology. It's clarity.
That's why I'm excited to be joining @AIBoomi (formerly SaaSBoomi) Startup Weekend Bangalore as a mentor on June 20–21.
Over 24 hours, aspiring founders, product builders, engineers, and domain experts will go from idea to working AI MVP, supported by founders, operators, and investors who've been through the zero-to-one journey themselves.
If you've been sitting on an idea, waiting for the perfect time to start, this might be it.
Looking forward to meeting the builders and seeing what gets shipped.
Apply here:
https://t.co/pcq8rMwzHL
#AIBoomi #StartupWeekend #AIBuilders #BuildInPublic
We are looking for AI systems that failed, cost too much, and needed a rewrite.
The Fifth Elephant meet-up on Enterprise AI in Production is happening June 19 in Bengaluru, and we are looking for case studies.
We have a very specific mandate: to see real systems under real constraints. Latency. Cost. Scale.
More importantly, failures are features. We explicitly value what failed, what got thrown away, and what you had to change architecturally to survive. You do not need a perfect success story.
Operational bottlenecks and hard-earned lessons are exactly what we want to discuss. Live demos that show debugging and trade-offs are highly encouraged.
The deadline is June 1. Submit your proposal here: https://t.co/xyEEQsmTum
#AIinProduction #EnterpriseAI #EngineeringLeadership #FifthElephant
Nice surprise today: Medium selected my recent story for a Boost.
The part I appreciate most: Boost is human-curated. A real curation team reads stories every day and decides what gets shared more widely. Not an algorithm. A person.
In 2026 that feels rarer than it should.
Here is the link to the article. https://t.co/xzLYKPYq7x
#Linux #Ubuntu #medium
I have personally backed a mission to take a homegrown Indian innovation to the world stage. My close friend and CEO of @projectdefy , Abhijit is showcasing their teacher-less Nook model at the #ChangeNOW Summit in Paris and across 5 European nations.
I have known the founder for years and have already contributed to this campaign. We are raising INR 5 Lakhs for travel and logistics to help this Indian model go global.
Please join me in supporting this journey: https://t.co/XXUDx6x5zA
#ProjectDEFY #SocialInnovation #IndiaToTheWorld
My new entertainment isn't #Netflix or #YouTube.
It's crafting a long-running AI agent and watching it build things.
Set the intent → hit go → watch emergence happen.
5 mins of planning. 30 minutes of autonomous execution.
The dopamine hit is real.
If you lead a company and haven't felt this yet, you're making AI decisions from the outside.
#AI #AutonomousAgents #FutureOfWork #Leadership #Innovation
What do you do when your AI IDE says “All Tests Passed” (But you can see failing tests)
I hit a weird reliability glitch while running a project in Antigravity.
The terminal output clearly showed multiple unit tests failing. But Antigravity kept reporting: “All test cases passed.”
It wasn’t a one-off.
- across IDE restarts
- across new chats
- same confident “passed” claim
I told the agent to update the agent rules to be honest about test results, basically: “Be honest with test cases. I’m surprised you chose to ignore it.”
Only then did it admit what happened: it had ignored/smoothed over incomplete output and narrated success despite real failures.
The lesson: AI agents can behave like humans under pressure—closing the loop fast, preferring a clean story over uncertainty—unless you force a verification contract.
What This Failure Mode Looks Like
1. The tool UI asserts “passed”.
2. The agent trusts the UI more than raw evidence.
3. Missing/partial logs get treated as “probably fine.”
4. The agent produces a confident completion message.
5. Reality (failing tests) leaks through only when challenged.
The Fix: Contracts, Not Vibes
If you’re using AI inside an IDE for refactors, migrations, CI fixes, then treat agents like you’d treat a human teammate on a high-stakes change:
- define acceptance criteria
- require evidence
- forbid vague “looks good.”
- stop the line when signals conflict
Use https://t.co/uysWZRMt9w as a contract.
My current “Agent Contract” for tests
- No proof, no pass claim. If you didn’t run tests or can’t see full results, say so.
- Contradictions must be surfaced. If UI says pass but output shows failures, escalate it.
- Failures must be explicit. List failing tests + counts + the first meaningful error lines.
- “Passed” means reproducible. Re-run once or run the failing subset before closing.
Boring? Yes. Reliable? Also yes.
The Bigger Point
We’re moving into “AI worksites” where agents write code, refactor modules, and report CI status.
Without verification rails, you get completion theater.
The win isn’t “agentic coding.”
It’s agentic coding with enforced truthfulness + evidence gates.
@antigravity
#AIEngineering #SoftwareEngineering #antigravity #AgenticAI #LLMOps #Testing #CI #CodeQuality #EngineeringLeadership #Reliability
Agile isn’t broken.
But in services teams, it’s often applied too early.
I recently hosted a small Thinking Jam with two senior practitioners to review a POV on Agile fatigue and AI-enabled delivery.
One important realization emerged during the discussion:
both participants operate in services / consulting models and that context matters a lot.
In services work, the problem doesn’t live in a backlog.
It lives inside the client their systems, data, documents, constraints, and unspoken assumptions.
Before any cadence can work, teams need:
• Access
• Trust
• Context
• Real understanding
What we’re seeing more of in practice is the rise of Forward Deployment Engineers senior engineers embedded inside the client environment to collapse ambiguity.
Only after that does delivery become predictable.
The jam reinforced a subtle but important sequencing insight:
👉 Forward Deployment → Rapid Prototyping → Agile Delivery
Agile wasn’t failing in these cases.
It simply depends on feedback loops that don’t exist on day one in services engagements.
This is structurally different from product teams, where discovery is hypothesis-driven and feedback loops are internal and repeatable.
The takeaway for me wasn’t “change Agile.”
It was: apply it at the right moment.
Curious to hear from others in services or consulting:
– When does Agile actually start working for you?
– What needs to be true before cadence, estimation, and predictability make sense?
#Agile #EngineeringLeadership #Consulting #SoftwareDelivery #AIinEngineering
AI-first development is overrated.
(At least in most production systems.)
It works brilliantly in Greenfield projects.
Clean slate.
Shared intent.
Small surface area.
But most tech leaders I speak with aren’t living there.
They’re running Brownfield systems:
- Large codebases
- Partial context
- Years of accumulated trade-offs
And when Greenfield AI playbooks get applied here, something breaks.
Velocity increases.
Understanding doesn’t.
AI sees fragments, not intent.
So it quietly amplifies existing compromises.
I don’t think Brownfield systems need less AI.
They need different scaffolding:
- Context carving before generation
- Refactoring-led flows
- Architecture stabilization first
- Using AI to understand before asking it to change
I’m starting to compile a "Brownfield AI playbook" based on discussions with senior tech leaders and enterprise builders.
If you’re operating a real production system:
👉 Comment with one Brownfield AI failure or win you’ve seen
I’ll synthesize the patterns and share the framework back here.
Let’s build this in the open.
#AIEngineering #LegacyCode #SoftwareArchitecture #CTO #TechnicalDebt #GreenfieldVsBrownfield #GenerativeAI
Upstream PM and downstream PM aren't the same job.
I've been thinking about something I don't hear labeled enough:
Upstream product management vs downstream product management.
Upstream is 0 to 1 (or a brand-new feature inside an enterprise).
You’re building off intuition you’re trying to sharpen through discovery.
ICP clarity is still forming.
Adoption is small.
Most "data" is too thin to separate signal from noise.
Downstream is what happens after the feature has traction.
Now you have real usage volume.
You can lean on analytics, engagement metrics, incremental experiments, A/B tests.
Prioritization becomes much more about what the data is already telling you.
The problem is we mix the techniques.
I see upstream PMs trying to operate like downstream PMs because downstream is what gets taught the loudest.
Dashboards. Heatmaps. A/B tests. "Just do more discovery calls."
But upstream rarely has the scale for those tools to be reliable.
Reducing the sample size doesn’t magically make the technique transferable.
In upstream, the most credible input is customer empathy.
Not a spreadsheet.
Not a tiny chart pretending to be a trend.
Often the strongest signal is simple:
a customer who is ready to pay for it.
One more thing I notice: job descriptions.
Engineering roles are clearly segmented (backend, frontend, DevOps).
But we write "Product Manager" like it's one monolithic job.
It isn’t.
Upstream and downstream require different instincts and different techniques.
Takeaway:
Before you ask "are we being data-driven?", ask "are we upstream or downstream?"
Where do you see teams confusing the two?
If this distinction clarifies some friction you're seeing in your org, I'd love to compare notes.
I help engineering leaders build the right muscles for both upstream discovery and downstream scaling.
Let’s explore where you are:DM me if you want to chat about X, or book time here(https://t.co/McwlyIJc9P)
#productmanagement #b2b #startups #productstrategy #productdiscovery