Sarang Bondre on 'Building AI Agents for Logistics'

How AI is transforming logistics with real-time orchestration, telemetry, and human-in-the-loop systems. A must-listen for CTOs and tech leaders.

# 🎧 Why You Should Listen

In this episode of Gravitas WINS Conversations, Joseph Jude interviews Sarang, Head of Engineering at Smile Digital and AI Labs, about how AI is transforming the logistics industry—not just in software, but in orchestrating real-world, physical operations like warehousing and transportation.

This conversation is a must-listen for engineering leaders, CTOs, and operations professionals who are curious about how to deploy AI in complex, physical systems.

It covers everything from telemetry data to orchestration engines, the role of human validation, AI-assisted SDLC, and team culture in AI adoption. If you're building or managing AI-first systems that interact with the real world, this episode offers hands-on insight and strategic direction.

# 🔑 5 Major Points

# 1. AI’s Role in Logistics and Intra-Operations

  • AI enables real-time orchestration between warehouses, vehicles, and delivery staff.
  • Reduces human coordination and streamlines operations with predictive systems.
  • Enhances customer experience by providing accurate delivery insights.

# 2. Telemetry as the Foundation for AI Decisions

  • Combines data from GPS, warehouse cameras, and IoT-enabled locks.
  • Acts as the “eyes and ears” of the system, feeding live context to AI agents.
  • Helps in tracking, routing, and setting correct customer expectations.

# 3. Engineering Culture and AI Adoption

  • Engineers are encouraged to treat AI as an assistant, not a competitor.
  • Tools like Cursor, Claude, and v0.dev are used with flexibility.
  • Sprint retros help decide the most effective tools through team feedback.

# 4. Testing and Validation of Probabilistic AI Systems

  • Early deployments involve human cross-validation of AI decisions.
  • Ground teams train models through lived operational feedback.
  • AI models are penalized/rewarded to improve through reinforcement learning.

# 5. Mindset and Leadership for AI Transformation

  • Focus on understanding problems deeply before building solutions.
  • AI should be used to explore nuances and dimensions of real-world challenges.
  • Leaders must give vision, not tasks—empowering teams to explore and learn.

# 💬 5 Memorable Quotes

  • “If I got even a one rupee for every time I heard ‘Where is my shipment?’, I’d be enjoying life on a beach.”
  • “Think of telemetry as the eyes and ears of the ecosystem.”
  • “You can’t win against AI. Use it as your assistant.”
  • “Backend engineers leaned toward Cursor, frontend folks preferred V0.dev. It happened naturally.”
  • “I'm looking for a time when orchestration will be so automated, two vehicles compete with each other for faster delivery.”

# Edited Transcript

Joseph:
Hello and welcome to Gravitas WINS Conversations. I'm your host, Joseph Jude.

Generative AI has mostly been applied to digital problems—writing text, generating images, and coding software. But what happens when you bring AI agents into the physical world, where trucks get stuck in traffic, warehouses run out of parking space, and even weather can disrupt operations? That’s where real complexity begins.

My guest today is Sarang, Head of Engineering at Smile Digital and AA Labs. His team is building AI agents not just to handle customer queries across email and WhatsApp, but also to optimize intra-logistics—managing real-time warehouse operations using telemetry, weather data, and predictive models.

It's a fascinating example of how AI can directly impact physical businesses. We’re going to discuss how to build AI-first systems for logistics, what makes the physical world different for AI, and what engineering leaders need to consider when deploying AI at scale.

Before we get into the interview, can I request you to subscribe to the podcast, write a review, and share it with your fellow business leaders?

Hello Sarang, welcome to the conversation.

Sarang:
Hello. Thank you for hosting me.

Joseph:
Let’s start with this. Tell me about Smile, your role in it, and what problems your team is solving with AI.

Sarang:
Smile is in the supply chain and logistics domain. We're also focused on adopting new development tools and software ecosystems. Specifically, my role is to enable businesses with actionable data. Along with my team, I'm solving problems around how shipments move, enhancing customer experience, and optimizing intra-logistics.

Joseph:
Why is logistics a particularly challenging but interesting domain for applying AI?

Sarang:
I wouldn’t say it’s hard, but it is definitely challenging. For example, when you order something from Amazon, you expect it to arrive quickly. But there’s a massive amount of coordination behind that—between humans, warehouses, shipments, and delivery expectations. That coordination is what makes it complex and interesting.

Joseph:
How do you handle that complexity?

Sarang:
Before AI, it was a sequential process: you place an order, it goes to the warehouse, stock is checked, then another warehouse is contacted if needed, a rider is assigned, packing is done, and then it's dispatched. Everything was managed manually and linearly.

With AI, we’ve removed many of those blockers. We now have an orchestration engine that coordinates across multiple systems—including humans—so the entire process is streamlined.

Joseph:
You talked about the "before AI" and "after AI" scenarios. We’ll dive deeper into that. But first, you mentioned how customers want deliveries to be immediate. They keep checking the status—I do that too. So I assume there are a lot of customer queries. How do you handle those? Is AI involved in managing them, especially since they often involve humans?

Sarang:
Absolutely. Every customer wants to know where their shipment is and when they’ll receive it. Honestly, if I got even one rupee for every time I heard that question, I’d be enjoying life on a beach.

Traditionally, when a customer care agent receives a query, they have to find the last known location of the shipment, identify the area or hub manager, contact them, understand the real situation—which could be delays due to traffic or other issues—and then give a polite but vague response like, “You’ll receive it in an hour.” But even the agent isn’t sure if that’s true.

Now, with telemetry and AI, we can pinpoint the exact location of the shipment, send messages automatically, and set the right expectations. This happens in real time—no more long wait times on calls. You can see where your package is and when it’s likely to arrive.

Joseph:
You mentioned telemetry. What kind of telemetry data are you collecting, and why is it crucial?

Sarang:
Telemetry is key, especially with AI. It includes data from web interfaces, GPS data from vehicles, camera feeds from the warehouse or delivery vehicles, and even smart locks with GPS that log access. Think of telemetry as the eyes and ears of our logistics ecosystem.

Joseph:
With all this data coming in, are you building your own models or using existing ones?

Sarang:
We’re not building from scratch. We take open-source models and fine-tune them on our logistics and supply chain data. This lets us experiment quickly while staying focused on our domain.

Joseph:
That’s a common pattern—start with a foundation model and train it on domain-specific data. Now, walk me through how this works in a real scenario. Say a customer places an order and wants to track it. How does telemetry come into play and how does the model respond?

Sarang:
Let’s say you place an order. First, we identify where the inventory is—could be in a nearby warehouse or a different city. Then, we assign a vehicle based on the size and shape of the item. For instance, you can’t deliver a sofa with a bike, or a small box with a truck.

We coordinate warehouse activities, assign the right rider, and notify you at each step. It’s just like what Swiggy does when you order food—you can see the rider moving on the map. That removes anxiety and reduces customer care calls.

Joseph:
So while AI plays a key role, customer psychology and UI design are equally important?

Sarang:
Exactly. As the tech lead, I don’t focus on UI decisions—that’s up to the product owners. But the experience differs depending on the category. On Amazon, you may just see "out for delivery." On Swiggy, you see the rider moving. That’s intentional design, based on user expectations.

Joseph:
Now let’s talk about testing. These AI models are probabilistic. How do you validate outputs in a logistics system where people expect deterministic answers?

Sarang:
That’s a great question. Testing AI in logistics is tricky. Initially, we deployed AI in controlled environments. For example, we asked our model to suggest optimal routes for shipments and cross-checked them with human experts. This helped us validate outputs and improve models.

You can’t afford mistakes in logistics—wrong temperature zones or routes can ruin shipments. So our ground teams help train and validate the AI. They're not just users—they're the real trainers of the models.

Joseph:
That brings up a concern I often hear—if humans are training models today, what happens when those humans are replaced? Who trains the AI then?

Sarang:
We don’t have to manually validate forever. AI models learn through reinforcement—they’re rewarded for correct answers and penalized for wrong ones. Over time, they improve and need less oversight.

Also, we’re not replacing humans—we’re evolving with AI. For example, if there are multiple routes to Goa, AI can factor in road conditions, weather, and delivery urgency to choose the best one. These are parameters humans may miss, but AI can optimize over time.

Joseph:
Let’s talk about the software development lifecycle (SDLC). Has that changed in AI-driven logistics? Are there new tools or processes?

Sarang:
Yes, absolutely. The biggest shift is a data-first mindset.

Initially, engineers felt like they were competing with AI. I told them, “You can’t win. Use AI as your assistant.” I gave them the freedom to use any tools they liked—Cursor, Claude, Lovable, whatever worked for them.

That changed everything. Now they create HLDs, LLDs, and test cases with AI. Earlier, they might miss edge cases. Now, they identify trade-offs and ask deeper questions. Development time has reduced by 70–80%, and the quality has improved.

Even while writing functions, they start with AI-generated base code, layer in tests, and iterate until it’s ready for production. Fewer bugs, better productivity.

Joseph:
Can you list some tools you use regularly?

Sarang:
Sure. For test cases, we use Cursor but also others. We’ve partnered with Typo App for security and vulnerability checks. We’re also experimenting with DroneHQ, Lovable, and tools to generate queries or build data models.

Joseph:
There are so many tools. How do you decide when to switch? For example, Cursor vs VS Code vs SF vs GitHub Copilot?

Sarang:
Great question. We experimented a lot. Initially, some used ChatGPT, others moved to Claude. Then we ran retros at the end of sprints and compared outcomes—time saved, clarity, learning.

Eventually, patterns emerged. Backend engineers leaned toward Cursor, frontend folks preferred V0.dev. Over the last few months, we’ve settled on a few tools and even subscribed to paid versions.

Joseph:
Looking ahead, what excites you about GenAI in logistics?

Sarang:
I’m excited about full orchestration—human and machine coordination where two vehicles might even compete for faster delivery. I envision a system with IMS, OMS, DMS, and zero chaos.

Joseph:
Do you think it’s achievable?

Sarang:
I think I’ll get there.

Joseph:
That would be a logistics utopia.

Now, what books, tutorials, or courses helped you learn AI quickly?

Sarang:
My biggest push came from my manager. He constantly challenged me to explore new things. I took a hands-on course by Raj Eker under the WithAI program. I’m also reading AI for Managers—it helps me balance tech and business. I follow experts on LinkedIn and Medium, including Antipatty.

Joseph:
Speaking of LinkedIn, I love your deep-dive posts on architecture. How do you stay consistent?

Sarang:
I plan in 3-month blocks—what I want to learn and what I want to achieve. Then I document what I’ve learned on LinkedIn so I don’t forget it. That method came from you—consume, produce, and engage.

Joseph:
Thank you. If a CTO or architect wants to start with AI today—not building models, but using AI—what would you recommend?

Sarang:
Don’t chase everything. Pick a clear business problem where you want impact. Start collecting data—good or bad. Use AI to understand the problem in depth. That journey will show you what to build and how.

Joseph:
I love that: use AI to understand the problem better. Fall in love with the problem, and everything else follows.

Sarang:
Exactly. That approach has worked well for me.

Joseph:
You mentioned your manager. Sounds like their influence was key. Can you share more about their management style and yours?

Sarang:
Leadership means being there when there’s a fire—walking the path. My manager gave me vision, not tasks. He stood by me, supported my learning, and inspired me to grow. That’s how I lead my teams too.

Joseph:
You’re lucky.

Sarang:
Yes, I am.

Joseph:
Alright, let’s end with a few rapid-fire questions.

What’s the kindest thing someone has done for you?

Sarang:
A friend’s father paid my computer course fee when I couldn’t afford it. It was outside my curriculum, but I really wanted to do it.

Joseph:
Who has shown the best leadership quality in your life?

Sarang:
Several managers who stood firm during crises. I try to follow their example.

Joseph:
What’s your definition of a good life?

Sarang:
Peace with your friends and family—and enjoying your ride.

Joseph:
Thank you, Sarang. You could’ve spent your Saturday morning anywhere, but you chose to be here. I truly appreciate it.

Sarang:
Thank you, Joseph. It’s an honor to be on your podcast. I’d love to return with more topics.

Joseph:
We’ll definitely plan that. Thank you again.

Sarang:
Thank you, sir.

Joseph:
I hope you enjoyed our conversation. Can I request you to share the key takeaways? If you liked this episode, please share the podcast with others.

Have a life of WINS.

# Connect with Sarang Bondre

# Connect with me

# Your feedback counts

Thank you for listening. If you enjoy the podcast, would you please leave a short review on Apple podcast or on YouTube? It takes less than 60 seconds, and it really makes a difference in finding this podcast. And it boosts my spirits. 😀

I would like to hear what resonated the most with you in this episode. Either email me or share it on social media and tag me. Thank you for your support.

Have a life of WINS.

Gravitas WINS Radio is produced by Joseph Jude. Be sure to subscribe on Apple Podcast, Spotify,YouTube, or the podcasting app of your choice.

Please check out other episodes of Gravitas WINS Radio Interviews categorized by theme. For easy consumption, they are categorized by themes. If you are a fellow podcaster, you might be interested in How I Podcast or Tools I use.

Published On:
Under: #aieconomy , #gwradio