Flora Liu is a software engineer at Notion building go-to-market systems that connect customer information, product usage, and sales decisions. Her work makes shared customer context usable by both people and agents, bringing engineering discipline to workflows that otherwise depend on scattered records, manual research, and repeated handoffs.
Her practice also encompasses accessible, human-centered product development and creative web projects, including Casa Kind, Floppy World, and Cloudie Drops. Projects such as Floppy World, featured on her personal site, belong to this creative side of her work alongside the operational systems she builds professionally.
At Notion, Liu moved from product growth into GTM engineering. That transition changed her understanding of the field: work she had once regarded as a marketing-operations problem became a distributed systems problem. Lifecycle messaging, product recommendations, onboarding, and sales automation had long addressed pieces of the customer journey. With a small cross-functional team spanning customer experience, revenue operations, product, engineering, and sales, she began bringing those pieces into a coordinated system.
The difficulty starts with a mismatch between the customer’s experience and the company’s organization. A customer can move between self-service adoption and sales assistance as part of one continuous relationship. Internally, marketing, sales, and customer operations may each hold different information and make separate decisions. At Notion, customer records were spread across Salesforce, Gong, Outreach, and other tools; product usage lived in Snowflake; consequential buying context accumulated in notes and meeting documents. An incorrect contact-to-account mapping could undermine a representative’s trust. A delayed integration could trigger an action based on stale information. A note saying that a customer’s champion had left—or that further contact was unwelcome—could matter more than any structured field.
A shared system for people and agents
Liu organizes the architecture around a know, decide, act, and learn loop: assemble trustworthy customer context, choose the next appropriate step, execute it safely, and feed the outcome into later decisions. Humans and agents work within that same loop, with different responsibilities.
Snowflake computes modeled, versioned entities for accounts, contacts, workspaces, eligibility, and facts, including ownership and timestamps. DynamoDB serves denormalized customer profiles that agents can query without joins. Generated research and summarized notes use the same identifiers, keeping structured records and unstructured context connected. Notion makes that information available to representatives in a workspace they already use, where they can investigate accounts, ask questions, and act without assembling the customer’s history across many tabs.
Several choices distinguish Liu’s approach:
- Agents prepare work and humans approve it. In sales-assisted workflows, agents research accounts, gather context, and draft recommendations while representatives judge the results and own the customer relationship. The team deliberately kept agents from communicating directly with customers and treated customer-submitted forms as untrusted input. Liu advocates shadowing effective representatives before automating their routines: encoding a weak process gives an agent a weak starting point.
- Customer signals coordinate the next action. A signal is an event important enough to change what should happen next, such as a customer reaching an AI usage limit, requesting sales assistance, or raising funding. A signal service checks eligibility, determines ownership, and creates a task for a person or agent. Centralized routing rules help prevent duplicate sends and coordinate communication across channels. When no signal calls for an action, a predictive engine recommends relevant product features through lifecycle emails and in-app nudges.
- Recoverable workflows make agent work practical. Liu’s research and drafting workflows use Temporal to handle retries and resume execution after failures across enrichment, web search, and generation calls. For cold outreach, agents research an account, generate three email drafts, and score and revise the strongest candidate before placing it in a representative’s task queue. After a sales call, an agent extracts buying context from a Gong transcript and prepares a grounded follow-up. Each language-model step is traced for evaluation. Prepared tasks also give newer representatives concrete examples of effective research and follow-up.
- Own the customer context; buy general infrastructure. Liu treats build-versus-buy as a decision made at each architectural layer. The team uses vendors for capabilities such as orchestration, enrichment, email, and CRM while retaining control of company-specific customer models and context. That information drives the decisions and must remain inspectable and debuggable. Keeping it readable by people and agents also reduces the risk that they develop incompatible understandings of the same customer.
The learning layer connects actions to their consequences. Decision logs associate each action with its outcome, and engagement history informs whether the system should continue a conversation, advance to another step, or change direction. Liu has reported promising preliminary results while emphasizing that the implementation remained in development.
Her longer-term aim is to make agents first-class operators, progressing from drafting toward action within safeguards. The foundation is shared information and understandable decisions: representatives can see the context behind a recommendation, apply their judgment, and learn from what happens next.
