Account Interface

Problem Statement

Sellers are assigned accounts to target through territory assignments, rules-based routing, or other methods. Depending on the seller, this can range from a handful of accounts to thousands — sometimes even a single large enterprise with dozens of subsidiaries across geographies.

Accounts are typically organized in a table format: each row represents an account, each column an attribute, and each cell a specific piece of information. To be effective, sellers must develop a deep understanding of all of their accounts — knowing when to engage, why, and with what message.

The challenge is twofold:

  • Scale: No individual seller can manually research thousands of accounts across dozens of attributes.

  • Personalization: Each seller values different information. Some prioritize financial signals, others expansion news, and others leadership changes.

The result is an overwhelming manual workload that prevents sellers from operating at their full potential.


Features

We designed Clever Columns (CCs) to help sellers research at scale with the flexibility to adapt to different selling strategies.

  • Configurable research: Sellers provide a prompt with parameters such as time recency, depth of research, or desired output format.

  • Parallel research execution: Research runs across all accounts simultaneously, producing short, inline answers with citations for trust and verifiability.

  • Filtering and sorting: Inline answers are stored at the cell level, making it easy to filter and prioritize accounts.

  • Multiple sales plays: Sellers can maintain different tabs with different research CCs for targeting specific account groups.

  • Dynamic updates: Alerts can be set on columns with rapidly changing information. Weekly refresh cycles update research so sellers always see the most recent signals.

  • Digest emails: Weekly insights are compiled into a digest delivered every Sunday, giving sellers a clear starting point for the week.


System Design and Implementation

Delivering this capability requires handling tens of thousands of concurrent research tasks per seller, and scaling reliably across all users. Our design leverages the Rox Task Framework for orchestration and execution.

Account Entity Resolution

Before research tasks can run, we must resolve accounts from a customer’s CRM into Rox’s canonical account graph. This is a non-trivial problem because CRM data often contains:

  • Duplicates: The same company represented multiple times with slight name variations.

  • Subsidiaries: A single parent company with dozens of child entities (e.g., regional branches or product-specific subsidiaries).

  • Ambiguity: Overlapping or missing data fields across CRM records.

To resolve this reliably, we employ a multi-layered approach:

  • Domain Matching: Web domains are a strong anchor for identity. Accounts are clustered when they share the same root domain, with heuristics for handling vanity domains and redirects.

  • Firmographic Signals: Attributes such as industry, headcount, geography, and revenue are used to disambiguate near-duplicates.

  • LLM-Based Entity Resolution: Large language models evaluate ambiguous cases by reasoning over names, descriptions, and firmographics to decide whether records represent the same company, a subsidiary, or distinct entities.

  • Parent-Child Graphing: Once resolved, accounts are linked into a hierarchical graph, ensuring sellers see the right entity — either the global parent or the relevant subsidiary — depending on their territory and sales strategy.

This resolution step ensures sellers don’t waste time on duplicates, while still surfacing meaningful subsidiaries when they matter.

Task Orchestration

  1. Massively Parallel Research

    • Each user may have thousands of accounts, each with dozens of CCs. This results in tens of thousands of tasks that must be scheduled, executed, and monitored.

    • Tasks include web search, data retrieval, reasoning via LLMs, and aggregation of signals.

  2. Priority-Based Task Handling

    • Tasks triggered by sellers in the UI are treated as higher priority than background refreshes or CRM syncs.

    • This ensures responsiveness when sellers are actively engaging with the product.

  3. Controlled Parallelism

    • Deep research tasks may involve chained LLM calls that vary in duration.

    • The framework parallelizes heavily to improve throughput but throttles execution to avoid rate limit violations and maintain reliability.

  4. Scalability and Reliability

    • The system auto-scales ECS workers based on queue depth and CPU utilization.

    • Retry policies with exponential backoff ensure resilience against transient errors.


Challenges and Solutions

Avoiding Extra Work

Running research on thousands of accounts is expensive, especially when LLMs are involved. To optimize efficiency:

  • Tab-based scoping: Research is run only on the subset of accounts within the active tab, not across the entire list.

  • Duplicate prevention:

    • Distributed execution and retry policies can lead to duplicate tasks.

    • We implemented Redis-based cell-level locks to enforce that each cell is processed only once, while still allowing full parallelism across different cells.

    • Executors check whether content has already been generated and marked completed before running. If so, they skip redundant work.

  • Time-ordered refreshes: Refresh cycles maintain strict ordering so that only the latest completed task is used, avoiding stale overwrites.

This ensures the system remains fast, cost-efficient, and reliable even under heavy workloads.


Why This Matters

Modern sellers face a paradox: they need to deeply understand their accounts, but they are responsible for too many to manually research. Clever Columns, powered by the Rox Task Framework, solve this by:

  • Scaling research across thousands of accounts without compromising on depth.

  • Personalizing workflows so each seller sees the information most relevant to them.

  • Maintaining trust with inline citations and transparent task logging.

  • Delivering timely updates via automated refreshes and weekly digests.

By combining flexibility, scalability, and reliability, we’ve built a system that turns overwhelming volumes of fragmented account data into prioritized, actionable insights. This empowers sellers to focus on what matters most: creating meaningful conversations and closing deals.

Last updated