Building the agentic engineering platform that knows ten years of d.light's system

• Category: Operate
• Industry: Off-grid solar / Pay-as-you-go consumer finance
• Service: Agentic AI platform / Platform engineering
• Client: d.light

d.light started in 2007 selling small solar lanterns. Today it is one of the world's largest providers of affordable solar home systems - lights, phone charging, TVs, refrigerators, cookstoves - sold on a pay-as-you-go plan to households that have never had grid electricity, and it has impacted more than 200 million lives. Behind every product is a platform that runs the whole business: mobile-money payments in every market, SMS delivery of the unlock tokens that keep a customer's lights on, and integrations with dozens of payment providers.

Symphony has worked with d.light on that platform since the beginning. In 2026 the two teams built something new on top of it: an internal agentic engineering platform, running on Claude Code in d.light's own infrastructure, that understands ten years of code, data and operations and helps the team detect, understand and fix issues in the live system. It is in production, with a human approving every change that touches the platform.

Problem

This is not a typical SaaS system. If the service is down, people are in the dark. Payments arrive through mobile-money providers that differ by country and sometimes by region; tokens go out over SMS providers that fail without warning; customers are often on 2G in places with no Wi-Fi. d.light's engineers design for the worst case first, because in this business the worst case is the base case.

Ten years of building for that reality produced a system too large for any one person to hold in their head. Developers and product managers could work on the platform for years and still not understand one part of it. When something broke at 10 p.m., whoever was on call had to read code, documentation and logs before they could even start on the fix - and every hour of that was an hour customers might be without light. The same gap slowed onboarding a new developer, scoping a new feature, and answering a product owner's question about how a payment flow actually behaves.

d.light wanted AI to carry that load. But the platform moves thousands of payment transactions a day, so the boundary was clear from the start: agents could investigate, explain and propose; they could not act on production without a human approving.

Solution

Symphony built a multi-agent system, running on Claude Code in d.light's own infrastructure, that understands the whole platform - ten years of code, data and operations - and puts that understanding in front of the engineers, product owners and support staff who keep the lights on. When an issue comes in, the agents detect it, raise the alarm and work through the code, data and logs to explain what is happening and propose a fix. Every agent runs with the least privilege it needs, and a developer or the business approves before anything changes.

Because Symphony had built and maintained the platform for years, the team knew where its real risks were and designed the agents around them - helping maintain and improve the system without touching the payment flow.

Technology: Claude Code, running in production on d.light's own infrastructure

Business Outcomes

• First response and root-cause analysis in about 10 minutes, down from a matter of hours: the agent starts working the moment a ticket is filed and has its analysis ready before an engineer opens it
• In production on d.light's own infrastructure with Claude Code as the core agent runtime, used by engineering, product and support, with human approval on every change

“The multi-agent system we built with Symphony understands everything we did in the last ten years building this software. Before, even our developers and product managers did not understand the full business logic behind the product - you could be implementing for years and still not understand one part of the system. Now, if there's an issue at 10 p.m., it helps us detect it, raise alarms and solve it. It's still a human-in-the-middle approach, because we're dealing with thousands of payment transactions a day, but it's a game changer for us.”

https://symphony.is/our-impact/case-studies/dlight-case-study