How Lendi Group rebuilt its mortgage platform for the agentic era

Lendi Group, the Australian home loan platform formed from the merger of Lendi and Aussie, has spent the past few years turning a traditional broking operation into an AI-native, end-to-end property and finance platform.

This story is part of the 2026 iTnews Cloud Covered report. Read it for free here.

Founded around 14 years ago to digitise what was a mostly manual broker experience, Lendi now integrates with more than 30 lenders plus its own white-label products, and supports a national network of around 1,300 mortgage brokers.

Its strategic expansion has gone hand in hand with a deep re-architecture of its cloud and data foundations to support a new generation of AI agents.

When CTO Devesh Maheshwari joined a little over two years ago, the brief from Lendi Group’s founders was to become an AI-first company.

“That meant we had to stop thinking in terms of one-off AI use cases and start designing production architecture that could safely scale AI across the business,” he told iTnews.

The Lendi–Aussie merger had left a patchwork of microservices, databases and integration patterns, which was fine for traditional workloads but a bottleneck for AI agents that need to reason over consistent, connected data.

Building an AI-ready cloud and data spine

Lendi Group now runs almost entirely on AWS, with 96–97 percent of workloads there and only minor use of other public cloud platforms. Rather than betting everything on a single, managed AI stack such as Bedrock, the team has built its own agent orchestration layer using the Amazon Elastic Kubernetes Service, and open source frameworks, with a strong focus on avoiding lock-in.

Maheshwari admits that Lendi was taking a bet in doing so.

“We borrowed the AWS principle around one-way door and two-way door decisions and were very deliberate about understanding what decisions we had to make and in applying the first principles of technology,” he explained.

A self-hosted LiteLLM proxy allows Lendi to route traffic to different model providers – Bedrock, Anthropic, Google’s Vertex AI or OpenAI – and even fail over between them within seconds when capacity is constrained. That “ultimate flexibility”, as Maheshwari calls it, is designed so the architecture can be deployed into new markets, without re-engineering core AI services.

On the data side, Maheshwari’s team has been consolidating and standardising what was previously a “fragmented” environment. Lendi has moved to MongoDB as an operational data layer that can represent complex lending and property objects closely aligned to business processes, while streaming into a Databricks lakehouse for analytics and AI.

Lendi’s data backbone now supports 14.5 million property records and enables creation of “digital twins” of each customer, used to model the impact of interest rate changes, equity shifts and future scenarios in near real time via the mobile app.

Agentic AI at scale – without the bill shock

Lendi’s AI initiatives range from customer-facing experiences to internal productivity tools. An agent-driven “Guardian” experience guides customers through the home loan process, orchestrating a swarm of agents behind the scenes to collect richer, more complete application data. Applications coming through this agentic funnel are processed up to 60 percent faster. On some metrics, Lendi is seeing up to 200% improvement in processing efficiency.

Brokers who previously handled three applications a day can now complete six or seven, because follow-ups and data gaps have been dramatically reduced. Importantly, brokers remain “in the loop” at the tail end, reviewing high-quality, AI-prepared files rather than being disintermediated.

Another flagship use case is contract-of-sale analysis. Customers or brokers can upload a complex, mixed-format contract and, within about a minute, receive a plain-language summary and a property report that outlines key risks and negotiating points. Maheshwari describes the initial temptation to solve this with GPU-heavy infrastructure, before a cost–benefit spike showed it would not scale economically for a free customer service.

The team went back to the drawing board and redesigned the architecture.

“With the contract-of-sale analysis we walked away from that design, spent three more weeks rethinking the architecture, and ended up with the same outcome at roughly one-tenth of the cost,” Maheshwari said.

Guardrails, risk and operating model

Because Lendi deals with sensitive financial data and regulated credit advice, security and compliance have had to evolve alongside AI. Traditional cloud controls around identity, access, misconfiguration, and data loss remain, but Maheshwari’s team has added finer-grained controls at the workflow and agent level. Agents inherit role-based permissions so that the same agent behaves differently depending on who invokes it, limiting potential damage from rogue access or misconfiguration. Centralised evaluation guardrails and an “LLM-as-a-judge” evaluation framework continuously score conversations in near real time.

“If it looks like the agent is breaching its boundaries, the guardrails kick in,” Maheshwari said. A risk team member is embedded into the AI design process, making “compliance by design” part of the operating rhythm rather than an after-the-fact audit.

Maheshwari is clear that the transformation is only “about a quarter to halfway” complete. But by treating AI enablement as an architecture and operating model problem, not just a tooling problem, Lendi Group has laid a cloud and data foundation that can support continuous AI experimentation without losing sight of cost, compliance or customer outcomes.

Maheshwari concluded: “Ultimately, this is not just a technology exercise. It’s an architecture and operating‑model challenge. Data quality, lineage, shared definitions – they’re all part of the day‑to‑day rhythm now, because without that, you can’t really be an AI‑native organisation.”