Solution

Insightly partnered with Caylent to evolve their internal prototype into a robust, enterprise-grade system. Rather than simply delivering features, Caylent co-architected the infrastructure, workflows, and interface patterns required to scale Insightly copilot into a strategic product offering.

Building a Task-Oriented Agentic Workflow

Technologies used: LangGraph, Insightly internal APIs, PostgreSQL (checkpointing & memory)

At the heart of Insightly copilot is an agentic system built using LangGraph, designed to handle real-world CRM workflows through natural language. Caylent redesigned the system to support complex, multi-step operations with built-in memory and decision logic. The agent was also enhanced to include checkpointing and rollback capabilities, enabling human-in-the-loop controls for sensitive actions such as deleting records. 

These workflows integrated fully with Insightly’s internal APIs, allowing the agent to perform real tasks like creating contacts, listing open opportunities, and summarizing support activity. This approach laid the groundwork for a CRM experience that feels conversational but remains grounded in reliable backend execution.

Embedding Intelligence Directly into the Insightly UI

Technologies used: React-based Insightly frontend, Native embedding (non-iframe), Page-aware state connectors

To ensure a seamless user experience, Caylent embedded Insightly copilot directly into the Insightly web application. The chat interface was designed to be page-aware and context-sensitive, so that users interacting with a specific contact or deal record could issue commands relevant to their current view. 

Rather than rely on iframes or external interfaces, Insightly copilot was integrated natively, enabling tight control over behavior and improved performance. This UI architecture also allows for future reuse across other applications or modules within the Unbounce ecosystem, improving extensibility.

Adding Memory and Personalization to Agent Behavior

Technologies used: PostgreSQL, LangGraph memory nodes

To deliver more personalized and coherent interactions, Caylent implemented a memory layer using PostgreSQL. This gave the agent the ability to retain multi-turn context, remember past actions within a session, and eventually support persistent user profiles. For example, a user can refer back to “the contact I just added” or undo an action with a follow-up message. 

These capabilities not only improve user experience, but also enable a more natural interaction flow that aligns with how people think and communicate. This memory system is designed to support long-term roadmap features such as user-level personalization and multi-session thread continuity.

Maintaining Agent Quality Through Automated Evaluation

Technologies used: DeepEval, G-Eval, CI/CD pipeline

As the technology evolved, Caylent prioritized maintaining a high standard of quality and reliability. The team implemented a full CI/CD-integrated evaluation framework using DeepEval and G-Eval. This system runs automated tests against a golden dataset of CRM queries curated by the Insightly product team. 

Each feature update is scored for correctness, coherence, and hallucination risk, helping the team detect regressions early and ensure consistent performance over time. These evaluations are run continuously as part of the development pipeline, providing both transparency and accountability for agent behavior.

Enabling Secure, Multi-Tenant Deployment at Scale

Technologies used: PostgreSQL with row-level security, Custom JWT-based auth with asymmetric encryption

To support Insightly’s growing customer base and enterprise requirements, Caylent implemented a secure, multi-tenant backend architecture. This design uses Postgres row-level security to enforce strict data separation between organizations, teams, and users. JWT-based asymmetric token authentication ensures that all interactions between the frontend and backend are encrypted, verifiable, and isolated. 

This architecture allows Insightly copilot to operate within highly regulated environments while maintaining performance and flexibility. It also supports role-based access control and org-level isolation, enabling Unbounce to serve customers with diverse and complex organizational structures.

Eliminating AI Vendor Dependence for Long-Term Flexibility

Technologies used: Amazon Bedrock (Claude 3.5 Sonnet)

As part of their strategic roadmap, Unbounce sought to move away from a dependency on OpenAI and transition to a more flexible and compliant foundation. Caylent facilitated a smooth migration to Anthropic Claude via Amazon Bedrock, ensuring compatibility with the existing LangGraph workflows while improving data residency and governance posture. This transition required careful validation of behavior, output consistency, and performance tuning — all of which were handled with minimal disruption to ongoing development.

Infrastructure as Code and Operational Observability

Technologies used: AWS CloudFormation, Amazon EKS, AWS Secrets Manager, AWS SSM, Amazon CloudWatch, Datadog, MLflow

To support reproducible deployments and internal handoff, Caylent aligned the system architecture with Unbounce’s existing AWS practices. This included the use of AWS CloudFormation for infrastructure-as-code, containerized deployment through Amazon EKS, and secure parameter and secret management via AWS Secrets Manager and AWS SSM.

Observability was handled through Datadog for primary monitoring and Amazon CloudWatch for supplemental insights. Caylent also leveraged MLflow for tracking tool trajectories and performance metrics during development. Together, these tools enabled a robust MLOps/LLMOps foundation that supported evaluation, deployment, and long-term scalability. These systems provided full visibility into agent behavior and performance, helping both teams maintain confidence as Insightly Copilot evolved toward production.