Validation is the cheapest place in your whole pipeline to catch that, and almost every team treats it as a checkbox at the end. We built Aries to fix it. In beta it took testing from weeks down to minutes.

Season 1 of AI Afterhours, Signals & Noise, is about how organizations cut through noise, find the signals that matter, and turn them into better decisions.

In Episode 6, API Assurance with Aries, I sat down with Varun Vemula, our CEO at AZTRA, to kick off the software and digital infrastructure arc. If you own APIs in production, you’ve lived this.

What we get into:

• The support ticket that started Aries, and why the team who got paged couldn’t fix it

• Why every tool out there checks whether an API responds, and almost none check whether it actually did anything

• Why two test cases can beat fifty, and how you know which ones to skip

• How much of Aries is actually AI. Varun says 30%, and not for writing the tests

• The customer in Amsterdam with an Ohio phone number, and why fake test data makes the whole run worthless

• Why AI-assisted code made this harder, not easier

“You can generate two test cases which 100% covers everything, rather than generate 50 test cases that are of no value.”

Varun Vemula, CEO, AZTRA

I looked up a Postman stat before we recorded. The average app has 28 to 50 APIs. That’s a couple years old so it’s probably low. And as endpoints grow, the ways they can contradict or fail grow even faster. It’s multiplicative.

It’s a bit ironic when you think about it. The whole point of wiring these endpoints together is to do things automatically. So why is the way we test them still somebody clicking buttons? There have never been more PRs or more incidents, and a lot of it is slop nobody checked. You build something, say it’s done, then go to use it and nothing works, all the pages are broken. Validation is the long pole in the tent now. It’s what stands between vibe coding and something that actually works on the other end.

Episode 7, Preventing API Failures at Scale, is next. We’re staying in the Aries world, but going from catching problems in validation to keeping them from happening in the first place.

Listen to the full episode below. And if you want to try this on your own API, show us your spec and we’ll tell you what we can discern from it and where the gaps are.

Thanks for tuning in. See you next episode.

Get more from Sean Fleming in the Substack app

Available for iOS and Android

Get the app

Season 1 of AI Afterhours, Signals & Noise, is about how organizations cut through noise, find the signals that matter, and turn them into better decisions.

Episode 5: Optimizing Production & Inventory closes out the manufacturing arc, with host Sean Fleming joined by Shashank Punuru, founder of Dovient, Harsha Varun, Lead AI Architect at AZTRA, and Varun Vemula, CEO of AZTRA, back after stepping away for his executive MBA.

Shashank opens by flipping the script. Episode 4: Manufacturing with Dovient focused on maintenance, but he makes the case that maintenance is usually the second thing a plant solves, not the first. A new operation starts clean with process, suppliers, raw materials, and inventory. Maintenance only shows up a year or two later. So production and inventory take the front seat, and the real job becomes planning for surprises, because the plant that runs exactly to plan doesn’t exist. You need a backup plan, and a backup plan for the backup plan.

Harsha lays out the systems picture. Planning isn’t one thing, it’s four layers, from long-range capacity down to the shop floor manager changing the schedule on the fly. The base systems run a bounded optimization and stop there. They can’t model the shock that lives outside their scope. His COVID example lands it: when a plant shuts down, no system can answer a question as simple as which other plant can make the same finished good, because every plant names things differently. That intelligence overlay is still missing almost everywhere.

Varun reframes the whole tradeoff. It isn’t service levels versus inventory cost. It’s an uncertainty problem. Nobody in the room has agreed how much uncertainty they’re willing to carry. The discipline is the CEO and CFO aligning on one uncertainty budget, then making sure that tolerance carries downstream to the salesperson and the planner. You can’t eliminate uncertainty. You can only minimize it. And anyone selling a 100% accurate forecast is selling fiction.

The takeaway from the episode: the plants that win are the ones that get the digitization and the signals right, then stop firefighting and start scaling. Get the foundation in place, and the surprises stop running the business.

Episode 6: API Security with Aries is next, moving off the factory floor and into the digital infrastructure that quietly runs behind everything. Different environment, same problem the season keeps circling. Something breaks, nobody notices, and by the time the business finds out, it’s expensive.

Season 1 of AI Afterhours, Signals & Noise, is about how organizations cut through noise, find the signals that matter, and turn them into better decisions. In Episode 4, Manufacturing with Dovient, I sat down with Shashank from Dovient and Harsha, who was back after missing Episode 3. Varun was away completing his executive MBA. Get it done, Varun.

Shashank opened with the frame that carried the whole conversation. Before AI, the shop floor runs on delayed visibility, static maintenance schedules, and coordination that happens by phone call. Teams react to what already happened. With AI, the same teams can finally act before the failure arrives. Not because the machine got smarter. Because the decision layer now has context.

Harsha made the invisible visible. The core problem is not the signal. It is that 80% of what governs the outcome never makes it into any system. It lives in shift logs, maintenance notes, and the memory of the technician who fixed this same machine three years ago and is not on shift tonight. Dovient’s approach is to digitize that unstructured knowledge, build the graph that connects events across time, and let background agents surface the pattern before the failure compounds.

Shashank’s closing thought is the one to take with you. Pick the one KPI where static planning is already failing. Start there. Measure it. Then expand.

Next up is Episode 5, still with Shashank and the Dovient team. Capacity decisions, lead time management, inventory positioning, and working capital trade-offs.

Shashank, the before and after frame you opened with is the clearest version of this problem I have heard explained. Hard to top that one. Harsha, good to have you back.

Thank you for tuning in. See you next episode.

Get more from Sean Fleming in the Substack app

Available for iOS and Android

Get the app

Season 1 of AI Afterhours, Signals & Noise, is about how organizations cut through noise, find the signals that matter, and turn them into better decisions. In Episode 1, Retail & CPG, I sat down with Bob, Harsha, and Varun to break down why stockouts, overstocks, and missed margin are usually not isolated failures.

They are the downstream consequence of weak signals, slow planning cycles, and planning systems that still assume stores behave the same when they do not. Bob made that real from the operator side. Harsha explained why forecast quality depends on better signals, simulation, and planner knowledge that usually never makes it into the system. Varun brought it back to the business. If you are still planning backward, you are probably bleeding margin forward.

Next week we get more specific with Episode 2, Margin Protection at Scale. Same category. Same pressure. Closer to the money.

Thank you for tuning in! See you next week.

Get more from Sean Fleming in the Substack app

Available for iOS and Android

Get the app