Building AI that behaves like a reliable colleague rather than an unpredictable experiment is currently one of the toughest problems in engineering. Most systems can be successful in a demo but fall apart in the unpredictable mess of real-world deployment.
For engineer Pranav Pawar, solving that gap has become the center of his work. Through roles spanning venture analytics, healthcare automation, and now marketing infrastructure at startup Kalos, he’s focused on the basic principle of making AI reliable and trustworthy above all else. His approach centers on building agents that don’t just perform once, but continue to work under pressure — systems designed to adapt, iterate, and deliver consistent results.
Why Most AI Fails Before It Scales
AI has reached nearly every industry headline, but far fewer industries have seen it work at scale. Most systems can perform very well in controlled demos, only to unravel in production when faced with the complexity of how teams typically work, which typically means dealing with less polished data, unpredictable user behavior, or the relentless demand for uptime.
Businesses need systems that they feel they can trust, meaning they should adapt to shifting regulations, handle complex data, and produce results that humans can verify. Yet too often, the workflow is divided into different AI agents, each in charge of one task — whether it’s creating text, categorizing data, or optimizing bids in isolation — while effectively forcing people to manually connect platforms that were supposed to save them time.
The result is AI that lacks a proper architecture that lets independent components communicate. This is the challenge Pranav Pawar has built his career around.
Pawar’s Early Engineering Work
Trained in Materials Engineering at IIT Madras, Pawar’s professional career started at Bain Capital Ventures, where he joined a small engineering group focused on machine-learning infrastructure for deal-sourcing analytics. Pawar built the foundational ML models and designed modular data pipelines that moved models from Jupyter notebooks into stable APIs, ensuring all sorts of teams could use them.
He first started working with generative AI when he joined ClarityCare, a healthcare startup tackling prior-authorization automation for insurers. Pawar built the backend from scratch, processing real medical records without breaching privacy and audit standards. Instead of trying to replace human reviewers, he introduced a verification layer that let nurses confirm AI-generated summaries of reports. The system was effectively a success, reducing time on manual work by about 15%.
These experiences reshaped how Pawar thought about AI. At Bain, he saw how models required constant fine-tuning to keep them running. At ClarityCare, that insight took on new weight, where accuracy was a necessity for real patients and clinicians depending on the results.
It made him see firsthand that reliability shouldn’t be thought of as a secondary goal, but as a key asset that guarantees AI can be properly useful in real-world settings. “Reliability is what separates a product from a project,” he says. “You learn fast that the hard part isn’t getting AI to think. It’s getting it to deliver, every single time.”
Inside Kalos: Automating Marketing Campaigns
This focus on reliability is what Pawar now applies as founding engineer at Kalos, a company building what they call an AI-agent-powered marketing platform built to simplify the most time-consuming parts of B2B advertising. The company’s goal is to turn fragmented marketing workflows into coordinated, automated loops.
The Kalos platform runs on a coordinated network of agents, each handling a distinct part of the advertising process. One analyzes a company’s sales calls, prospect emails, and CRM notes to better understand what drives conversions; another designs full ads from text to images; a third manages the audience, sorting through various criteria to identify high-value targets; a fourth optimizes campaigns by adjusting bids and running the strongest ads; and a final agent evaluates campaign performance, matching up metrics with CRM opportunities and pipeline to show ROI.
Together, they turn what was once a manual process into a data-driven feedback loop of research, ad execution, and maintenance.
Beneath that interface lies the problem Pawar spends most of his time solving: orchestration. Each agent depends on the output of another, and one error can cascade across the entire system. His job is to ensure those dependencies resolve cleanly, which means: refactoring messy code, making sure logic is consistent across all agents, and making the product resilient enough to serve hundreds of customers. “Only a small part of AI is about the models,” he says. “The rest is the engineering that makes it reliable and scalable.”
Lessons On Building Lasting AI
Pranav Pawar’s work offers a few lessons drawn directly from the systems he’s built on how to make sure AI is reliable, specific, and usable in production.
- Address the pain point first: Every project Pawar has led began with a concrete bottleneck: manual prior authorization in healthcare, fragmented analytics at Bain, repetitive ad management at Kalos — each one needing solutions tackling specific problems.
- Keep humans in the loop: At ClarityCare, he saw that automation succeeds only when people stay part of the process. Nurses reviewed AI-generated summaries instead of being replaced by them, creating a system that strengthened trust.
- Separate prototypes from products: At Bain, Pawar realized taking a model live meant constant rebuilding and oversight. Reliability came from process and discipline, an understanding that later shaped how he approached scaling Kalos.
- Build around business metrics: Whether optimizing insurance workflows or marketing performance, Pawar focused on factors like throughput, efficiency, and conversion to measure the success of his models.
- Design for the long haul: At Kalos, each agent is built to continually learn from ongoing campaign data, which is crucial for the system to stay effective as markets and customer behavior shift over time.
These notions define the throughline of Pawar’s career. He doesn’t treat AI as a one-size-fits-all solution but as applied software built to solve specific problems in ways that last. “AI engineering isn’t about hype,” he says. “It’s about solving real problems reliably and at scale.”