Lasse Hyldahl Jensen

I joined Avido as its Founding Engineer and third co-founder, after working with Avido’s CEO in Lunar’s AI Colony. We help banks test, monitor, and audit the AI they put in front of customers. I’ve taken our product from a vibe-coded Next.js MVP to something banks run in production, onboarded our two engineers, and I run regular sessions on building and evaluating AI agents.

My main work is the evaluation platform, and it is largely a system of AI agents working together. Some agents simulate real end-user conversations against a customer’s AI; others grade the answers against evals like faithfulness and context recall. I built the evaluation frameworks and metrics behind that grading, with the error analysis and human-in-the-loop review that keep them honest. The quickstart I built adds more agents on top of clustering and vector search to turn a customer’s raw data dump into clean, usable test sets, and customers can stream their own production conversations in over OpenTelemetry and reuse them as test material. It runs on the Vercel AI SDK over OpenAI and Azure OpenAI, with Postgres and pgvector for the semantic search.

Alongside the platform I built the AI Core in Python and FastAPI, a configurable conversational-AI engine customers deploy to build their own assistants on. It runs the LLM inference in production with tools wired in through MCP, moderation as a guardrail, RPM and TPM rate limiting to stay inside one bank’s AI gateway, and every request traced back for evaluation. I own the rest of the product end to end too, from the Next.js frontend and Node.js (Hono) backend down to the infrastructure on Azure Kubernetes, with Pulumi, GitOps, CI, and Grafana for observability.

Avido is now used by companies like Skatteguiden, Sydbank, Kamina, and ABN AMRO. To get through one bank’s procurement we installed the software inside their own private cloud, shipping the same container images whether we host it or they do. Security is held to the standard banks expect, with passwordless Azure Workload Identity and zero-trust networking between services, and our SOC 2 penetration test came back with zero findings.

In early 2024, two colleagues and I started Lunar’s AI Colony, a small team set up to find out how far generative AI could go across the bank’s products and internal work. The brief was simple: get hands-on and show value quickly.

We spent the first month on internal tooling, an assistant for staff and a support chatbot wired into Lunar’s knowledge base. That convinced us, and the wider organisation, to put AI in front of customers, so we launched the customer-facing assistant in April 2024 and built a full support system on top once people trusted it.

The support system connects to Intercom through its REST API and answers customers in seconds, where a reply used to take up to three days. It resolves 85% of Lunar’s support cases today; for the rest, it gathers what a human will need and hands the case over.

Between them, the assistant and the support system handle over 40,000 customer interactions a month, and beyond support we put AI to work elsewhere in the bank too, including KYC and adverse media screening and a bill scanner that read payment details off uploaded invoices. It all ran on Lunar’s stack: Go microservices behind OpenAPI and GraphQL APIs, with retrieval over PostgreSQL and pgvector.

I also built Lunar’s internal GenAI platform, so other teams could reach LLMs through our APIs and build their own RAG pipelines on top, without standing up the plumbing themselves. On the data side, I built a pipeline that synced data from our warehouse back into production, so users could compare their own spending with others in their cohort.

I built backend microservices in Go, mostly services that exposed APIs to Lunar’s app and internal tools through OpenAPI and GraphQL.

As an intern I built an anti-corruption layer in front of one of Lunar’s partner integrations. API calls had been forwarded straight to the partner, which added latency and left little control over response times; the layer fixed that and let us move to event sourcing across the affected microservices. Those endpoints were hit every time one of Lunar’s million users opened the app, and we shipped changes to them continuously, sometimes several times a day.

Full-time, I took over Lunar’s card management system. It dated back to the 2019 banking licence and the rushed 2020 launch, so it carried real technical debt. I reworked the architecture that was holding back scalability and built an integration with Nets, Lunar’s card issuer, to keep the two systems in sync and reconcile any discrepancies automatically.

For my Bachelor’s project, I collaborated with IBM Denmark and Smukfest to develop the backend for Smukfest’s new mobile app. The backend was designed for scalability and maintainability, and it integrates various Content Management Systems (CMS) while enabling push notifications to all devices via Expo.

The app successfully launched for Smukfest 2023, supporting over 50,000 attendees, and was built to keep running even when parts of the backend failed.