From a conversation to a working business flow

A business conversation can involve a customer, a job, an invoice, a payment record and a document to send. Making those actions useful requires the web experience, agent behavior and backend services to agree on what has happened and what should happen next.

At New Media Lab, I worked across this conversational product and its growing web and creator-portfolio capabilities. My contribution combined hands-on, AI-assisted engineering with product decisions, integration review and coordination across the product team.

I helped turn the job, invoice, confirmation, cancellation and payment-record journeys into explicit product flows. I worked on web onboarding and plan behavior, editable messaging Flows and creator-portfolio generation, access, refresh and previews.

As the product expanded, I aligned responsibilities across its chat, portal, portfolio and identity services. This included API contracts, Keycloak integration and subscription behavior, so changes could be implemented and reviewed across the relevant services.

I refactored context and task handling to address conversational continuity, and worked on the master-agent flow and per-user behavior. The engineering questions included how a conversation retained the right task context and how to investigate a response that did not match the intended workflow.

I managed versioned Langfuse prompts and checked prompt/source compatibility. Traces supported behavior investigation and estimates of LLM workflow cost.

I coordinated multi-service releases and worked on portfolio access, workspace recovery and PDF improvements. Subscription work included plan integration and fixed-term redemption behavior. I also contributed to official Instagram OAuth and portfolio-generation preparation, including staging checks and reviewer journeys.

I introduced PostHog tracking and masked replay to help evaluate product behavior. Event verification and raw-IP removal were part of that work; the result was better instrumentation for evaluation, with no conversion-uplift claim.

I combined automated regression with selected live journeys from a WhatsApp job through invoice/PDF delivery, payment records and reconciliation. I also established local pre-PR validation expectations and coordinated fixes across frontend and backend changes.

The records through October 2026 include scoped production releases and recovery work. Instagram review preparation, live redemption acceptance and parts of event matching remained separate work in progress at that point. My role covered product decisions, implementation and coordination within a wider team.