Tarun Upaday.

Essays

Essays on AI agents and the systems around them.

Where agents help, where they fail, what has to sit around the model, and what building companies taught me about all three.

  • An AI task is not done until the world has changed

    I moved my working life onto AI agents and logged every failure. Each one traced to the same cause: the to-do item is the wrong primitive for a machine.

    Aug 16, 2026 · Thinking in Public

  • PillWatch: 99 percent accurate, wrong every fifth box

    Four high schoolers, a $60 Raspberry Pi, and the compounding math that turns 99 percent per cell into an error every fifth box.

    Aug 16, 2026 · Thinking in Public

  • 1,029 cells: what small, honest data teaches about AI

    The PillWatch team labels its own data, splits it by photograph, and hunts its rare cases deliberately. That discipline taught more than model tuning did.

    Aug 16, 2026 · Thinking in Public

  • The LLM call you can delete

    631 replayed decisions, six written rules, zero disagreements. The model I was paying to route my email lost to the policy file it was reading.

    Aug 16, 2026 · Thinking in Public

  • Pivot gates

    How founders can distinguish a slow strategy from a broken premise before sunk cost makes the decision for them.

    Mar 31, 2026 · Building Routespring

  • Routing optimization isn't a shortest-path problem

    The hard part of flight routing was never finding the cheapest itinerary. It's that 'cheapest' and 'best' are different questions, and one of them is a hard rule while the other is a matter of taste.

    Jan 19, 2026 · Thinking in Public

  • The first question in any outage: whose system is it?

    Most production failures in integrated systems aren't yours or theirs. They're the handoff between the two, and that's exactly where nobody looks first.

    Dec 8, 2025 · Operating Notes

  • Build versus buy is really a question about who owns the interface

    The standard framing, comparing the cost of building to the cost of buying, misses the question that actually decides it: does the vendor's product expose the interface you need, or just the outcome?

    Nov 17, 2025 · Operating Notes

  • What successful exits do and do not teach you

    An exit gives a founder credibility, not certainty. What building and selling companies actually teaches, and why a previous exit can quietly become a liability.

    Nov 3, 2025 · Operating Notes

  • How enterprise buyers evaluate operational risk

    Founders assume buyers are evaluating the product. They're also evaluating what happens when it fails. Why an inferior incumbent keeps winning, and how to sell controlled change instead of pure improvement.

    Oct 27, 2025 · Operating Notes

  • When services are a wedge, and when they become a trap

    Services can be the only credible way into a complex industry, or the thing that caps your margins forever. The distinction isn't whether humans are involved. It's whether their effort compounds.

    Oct 20, 2025 · Operating Notes

  • Why enterprise software fails at exception handling

    Enterprise software is designed around states. Real operations happen between them. Why systems fail at the moment users most need help, and what good exception handling requires.

    Oct 13, 2025 · Operating Notes

  • Manual minutes per transaction is better than "AI adoption"

    "AI adoption" is one of the least useful metrics in software. Manual minutes per transaction is hard to fake, captures hidden labor, and separates adopting AI from actually improving operations.

    Oct 6, 2025 · Operating Notes

  • Knowing whether a workflow is actually automatable

    Every workflow looks automatable from a distance. Up close it's a sequence of judgments, not clicks. Five questions to ask before you automate anything, and why the process that works 80 percent of the time is the most dangerous kind.

    Sep 29, 2025 · Operating Notes

  • Building when one customer dominates the business

    Concentration isn't the real risk. Failing to convert a dominant customer into reusable product is. The metric that matters isn't percentage of revenue. It's percentage of the business that could transfer to another customer.

    Sep 22, 2025 · Operating Notes

  • Company-Building Under Real Constraints

    A seven-part series on building operational software companies: customer concentration, automation, exception handling, services, enterprise risk, and what exits do and don't teach. Written from the operator's chair, not the slide deck.

    Sep 15, 2025 · Operating Notes

Archive

The original tarun.me blog, from 2009.

  • Why web designers are over-rated

    Wired did a competition inviting designers to improve craigslist .

    Sep 28, 2009 · Archive

  • when good people resign (to join you)

    You make an offer to an A+ player at a big company, they accept — and then their employer rolls out the red carpet to keep them. Mark Suster calls it the counter-offer, and it's an all too familiar phenomenon worth planning for before you extend the offer.

    Nov 18, 2009 · Archive

  • 2 things every good startup needs

    Robert Scoble has a great article today on his take on mistakes startups make .

    Nov 21, 2009 · Archive