Dayara Founder on Agent Orchestration, Consumer Trust, and Building Sustainable Travel by Default

Supratim Dey is the founder and head of product and sustainability at Dayara, an AI travel platform built on a multi-model agentic orchestration stack. Before founding the company, he advised global enterprises on climate and sustainability strategy at Boston Consulting Group, and earlier earned a PhD in planetary science from UC Davis, where he held a NASA fellowship analyzing meteorites to reconstruct the formation of the early solar system. We spoke with him about what separates a genuine AI agent from a chatbot with plugins, how experimental science shaped the way Dayara validates model outputs, and why his team builds sustainability directly into the recommendation engine instead of leaving it as a checkbox at checkout. 

Travel is one of the most complex consumer categories for AI to handle, with pricing, logistics, and preferences all shifting in real time. What made you decide an agentic approach was the right architecture for it? 

Traditional travel platforms were built as information lookup engines. They solved access to static data or real-time inventories and pricing (hotel directories, flight schedules, attraction lists) but left the heavy cognitive load of planning the trip entirely to the consumer. A single vacation requires balancing dozens of interdependent variables in real time: dates, budget limits, transit friction, dietary constraints, opening hours, geographic clustering, and individual preferences. 

When you change one variable, say, it’s supposed to rain the day you planned an outdoor activity or shifting a hotel location, it causes a cascading failure across the entire itinerary. Traditional search engines and standard LLM chatbots cannot handle this because they are inherently linear and reactive. 

An agentic architecture is fundamentally different because it treats travel planning as a dynamic, high-dimensional constraint-satisfaction problem. Human travel agents succeed because they can reason and adapt with changing constraints. At Dayara, we are building an agentic framework to replicate and scale that exact multi-step reasoning process. The agent doesn’t just answer queries; it maintainspersistent state, evaluates trade-offs across complex logistics, and dynamically replans your trip when real-world conditions change. For example, if sudden rain cancels an afternoon outdoor tour, Dayaradoesn’t just alert the traveler; it autonomously suggests an indoor cultural alternative nearby, adjusts downstream dinner timing, and recalibrates transit routes while preserving all dietary and pacing preferences. For a domain as fluid and complex as travel, anything less than a fully agentic system breaks down immediately. 

Dayara runs on a multi-model orchestration stack rather than a single model. Walk us through how that works in practice and why routing across models matters for a consumer product. 

Relying on a single frontier model to power an enterprise consumer product is an architectural antipattern. It forces you into severe trade-offs between latency, compute cost, and reasoning quality. In a real-time consumer product like Dayara, different micro-tasks within the user journey require radically different model capabilities, and this directly impacts both user experience and user trust. If an AI agent takes ten seconds of “thinking” time just to answer a simple question about local tipping customs, the user experience feels sluggish and frustrating. Conversely, if an agent instantly spits out an unverified itinerary with impossible transit connections, user trust is destroyed permanently. Delivering a successful consumer product requires balancing a delightful user experience with rock-solid logistical accuracy. 

Dayara’s orchestration stack operates as an intelligent routing layer that dispatches specialized tasks to the optimal model based on intent, complexity, latency budgets, and cost metrics. When a traveler asks for quick contextual details, like understanding local culture, weather checks, or simple spatial lookups, Dayara routes the request to lightweight, fast-inference models that deliver sub-second response times without burning unnecessary compute. For complex, multi-constraint schedule optimization, where transit times must match traffic patterns, opening hours align, and a sunset viewpoint isn’taccidentally scheduled at noon, we route the task to frontier deep-reasoning models capable of handling complex logic chains. Finally, before any plan is presented to the user, the output passes through deterministic verification layers and specialized evaluation code that audit the schedule against hard logical rules. This deterministic layer ensures that the itinerary matches real-world temporal and spatial constraints, such as verifying museum opening hours or catching impossible transit windows like scheduling a 1:00 PM tour 30 miles away from a 12:30 PM lunch. 

Routing across specialized models isn’t just a backend optimization trick; it is the foundational architecture for building consumer user experience and long-term user trust. By matching each task to the right model capability and deterministic guardrail, Dayara delivers the fluid responsiveness users expect while optimizing for our token costs. 

There’s a lot of debate about the difference between a genuine AI agent and a chatbot with plugins. Where do you draw that line, and how does Dayara clear it? 

I draw the line at three fundamental capabilities: autonomy, state management, and resilient error recovery. Most tools labeled as “agents” today are really just chatbots with plugins, which are reactive prompt wrappers that pick an API tool, execute a single linear call, and format the output. They have no persistent memory of execution failures, no capacity for multi-step replanning, and crumble the moment an API returns unexpected parameters or when competing user constraints clash. 

A genuine AI agent, by contrast, is a stateful, goal-directed system. Where Dayara clears that bar is in how our orchestration engine handles autonomy and continuous feedback loops. 

Dayara builds a persistent understanding of each traveler through an onboarding profile and travel persona quiz. When a user inputs their trip destination, dates, and budget, Dayara surfaces tailored activity and experience options for the user to select from. The agentic engine then takes those selected choices, combines them with the user’s underlying persona preferences, and programmatically orchestrates a complete, seamless itinerary, while dynamically weaving in complementary activities, meals, transit links, and timing buffers. 

If an activity API returns zero availability, a venue is closed for private events, or the overall cost falls significantly outside the user’s selected budget range, a chatbot either fails to identify it or returns an error message. We have found that finding engaging things to do within a strict budget is a massive cognitive load and barrier to travel for many people. Dayara’s agentic architecture detects budget and logistical constraint violations internally, initiates an automated re-planning loop, finds alternative activities that fit the price parameters, and resolves the conflict before the user ever sees it. 

The system maintains continuous context across the entire trip. If a traveler chooses to modify an activity, swap plans between days, skip a tour, or if weather results in cancellations Dayara re-adjusts downstream timing, dining reservations, and transit logistics to keep the rest of the itinerary coherent and intact. 

Sustainability is often bolted onto travel products as an afterthought, like a carbon offset checkbox at checkout. How are you building it into the product’s core logic instead?  

Carbon offsets at checkout are a psychological band-aid, not a structural solution; they ask consumers to pay extra to fix emissions after the damage has already been planned into the itinerary. This dynamic drives the “say-do gap” that Booking.com has documented in its annual sustainable travel research, where a large majority of global travelers say they want to travel sustainably, yet only a small fraction actually do. Most travelers simply don’t know where to start, or they are discouraged by the perception of a “green premium”, the belief that eco-friendly choices are inherently more expensive. 

To tackle this effectively, it helps to understand our specific focus area. Flights tend to dominate the conversation about travel emissions, but a substantial share of any trip’s footprint comes from decisions made after arrival: how travelers get around, where they stay, what they eat, and which local operators they book with. Dayara works on that side of the equation, because that is where choice architecture can immediately reshape the experience for a traveler. 

At Dayara, we are proving that sustainable travel doesn’t require a price premium. Many of the most sustainable experiences, from local walking tours and farm-to-table dining to multi-modal regional transit, are actually very budget-friendly. We eliminate the cognitive friction entirely by making sustainability effortless: it is baked directly into the trip plan itself. The traveler doesn’t have to do any heavy cognitive lifting, pay extra fees, or spend hours researching. Dayara’s engine optimizes for low environmental impact by default while ensuring the trip remains fun, memorable, and personalized to the traveler’s exact preferences. 

At an architectural level, we integrate sustainability directly into the multi-objective optimization function of our recommendation engine: 

Ground Transit Optimization: Rather than defaulting to private rideshares or gas vehicles, the system programmatically prioritizes walkability, high-speed rail, public transit networks, or electric fleet transfers where time trade-offs are negligible. 

Hyper-Local Ecosystem Integration: The system evaluates supplier data to surface sustainable accommodations and local operators that genuinely minimize environmental footprints, weaving them organically into the itinerary. 

Frictionless Defaults: By making the most sustainable choice the default, highest-value option presented to the user, we lower carbon intensity by design without forcing the user to compromise onexperience or manually calculate environmental trade-offs.  

You spent time at BCG working on climate and sustainability strategy for large enterprises. What did that teach you about where AI can realistically move the needle on emissions, and where the hype outruns the data? 

My time at BCG advising corporate leadership on climate and sustainability strategy revealed a fundamental truth: sustainability initiatives fail when they rely on static industry averages, top-down estimations, and PR-driven metrics rather than granular, transaction-level operational data. I saw this firsthand while engineering Scope 3 emissions reduction roadmaps across global supply chains and evaluating clean energy strategies for energy-intensive AI data center deployments. 

In both cases, high-level pledges faltered during implementation because legacy systems relied on broad spend-based estimates and retrospective carbon accounting. In the climate-tech space, hype outruns data whenever companies treat generative AI as a silver bullet, expecting a standalone LLM to solve sustainability by spitting out broad recommendations or automated offset purchases. Asking a generic LLM for advice yields superficial statements like “taking the train is greener than flying”, which carry zero operational utility for an enterprise or a consumer. 

Where AI genuinely moves the needle, and what directly inspired how I architected Dayara, is in granular, real-time choice architecture and operational data integration: 

Replacing Spend-Based Estimates with Transaction-Level Activity Data: At BCG, I learned that true decarbonization requires moving away from top-down averages and mapping raw, granular activity metrics to verified emissions databases. At Dayara, I applied this exact principle to consumer travel: instead of giving users generic “carbon scores,” our engine parses real-time operational parameters—like specific transit modes, local supplier practices, and spatial routing—to calculate and optimize actual emissions at the point of recommendation. 

Automating Decision-Making at the Point of Selection: Whether you are shifting compute workloads to align with grid carbon intensity or building a consumer trip plan, behavioral change only happens when low-emissions options are presented as frictionless defaults. Dayara uses AI to model traveler preference curves so that the most sustainable option is woven into the itinerary as the most attractive, convenient, and cost-effective choice. 

BCG taught me that to make a real environmental impact, software must bridge the gap between climate science, transaction-level data processing, and effortless user experience. That exact strategy is what underpins Dayara’s recommendation engine today. 

Before industry, you earned a PhD in planetary science and held a NASA fellowship. How does that scientific training show up in how you build and validate AI systems today? 

My PhD research at UC Davis and my NASA fellowship focused on planetary science and isotope cosmochemistry, essentially analyzing ancient meteorites to reconstruct how our solar system formed billions of years ago. In cosmochemistry, you are working with incredibly sparse, noisy datasets where a tiny error can completely throw off your conclusions. It instills a relentless discipline: you learn never to take a plausible-sounding theory at face value, and you never trust a result without testing it against rigorous controls. 

When you transition into AI engineering, that mindset becomes an invaluable asset. Large language models are non-deterministic engines; they generate answers with absolute confidence, even when they are completely wrong. Coming from an experimental science background, I don’t view an LLM output as a final answer. I view it as an unverified hypothesis that has to earn its way into the product. Just as we used high-precision mass spectrometers and strict laboratory controls to verify isotopic data from the rarest meteorites, at Dayara we gate every AI output behind multi-layered testing pipelines, synthetic edge-case benchmarks, and deterministic verification code. That scientific rigor ensures we look past surface-level AI hype and focus entirely on building rock-solid agentic architecture that earns and keeps consumer trust. 

Consumer trust is a recurring problem for AI products, especially ones making recommendations with real money attached. How do you engineer for accuracy and reliability when a hallucinated flight time has real consequences? 

We achieve this by maintaining a strict separation of concerns within our architecture. Non-deterministic LLMs are used solely for intent parsing, natural language translation, and creative synthesis. They are never allowed to act as the single source of truth for factual, time-sensitive, or financial data. Instead, specialized sub-agents handle discrete tasks like experience selection, route planning, and timing orchestration, while a dedicated evaluator agent continuously monitors their output. This evaluator audits proposed plans against hard deterministic guardrails and queries live, audited APIs, such as activity booking feeds, mapping routing engines, and verified business rule engines, to confirm availability, opening hours, and exact timing. If the evaluator agent or deterministic layer detects any discrepancy between a generated plan and real-world ground truth, it automatically triggers a targeted re-prompt loop back to the orchestrator and sub-agents, forcing them to recalculate using corrected parameters long before the itinerary ever reaches the user interface. 

By gating non-deterministic AI generation behind dedicated monitoring agents and strict, deterministic validation pipelines, we ensure that every real-money recommendation Dayara makes is concrete, accurate, and dependable. 

Where do you see agent orchestration going over the next few years, and what will separate the teams that get it right from the ones that don’t? 

We are rapidly moving past the era of simple API wrappers and basic single-prompt travel tools. Over the next few years, agent orchestration in complex consumer verticals won’t just be about prompt chaining; it will evolve into specialized, domain-specific operating systems. The teams that succeed will be those that look beyond raw foundation model updates and build deeply integrated, domain-specific architectures that remain resilient as real-world parameters shift. At Dayara, we built our multi-model orchestration stack around this exact realization. Generic wrappers offer no defensibility and poor reliability when a traveler’s budget or schedule conflicts with real-world inventory. 

The winners will be defined by how deeply they anchor their agentic architecture in domain-specific problem spaces. At Dayara, that requires key technical pillars: engineering a proprietary multi-model router that dispatches specialized micro-tasks based on latency and reasoning requirements, merging generative personalization with dedicated evaluator agents and deterministic verification layers to eliminate logistical hallucinations, and building resilient error-recovery loops that autonomously handle mid-trip changes, budget mismatches, and schedule conflicts. Crucially, Dayara is also constantly generating data from explicit user choices and feedback, continuously refining our understanding of traveler preferences so every iteration delivers increasingly tailored recommendations and highly personalized itineraries. That architectural depth and preference-learning flywheel are what create an enduring consumer product, and that is where the future of agent orchestration lies.