Build an account intelligence console for a B2B sales team — a "battle-card" research tool reps use to research target companies with an LLM, work the results as pipeline, draft outreach, and manage their book of accounts. Entry point: A browser-facing page (SSO-gated) where a rep pastes in company names to research. A separate admin-only dashboard shows system health and spend. Everything besides the two pages is an API the UI calls. Core flow: Intake & queue. Reps submit companies; requests are queued and worked by a background queue processor (a scheduled/cron job claims queued accounts and dispatches worker runs) with retry, backoff, and dead-lettering for stuck jobs. Research. A worker step calls an LLM (e.g. Claude, or your model of choice) with live web search grounded in your own positioning/case-study docs, and writes a structured dossier per account per owner. CRM enrichment. After research, pull the matching account/contact/lead record from your CRM (e.g. Salesforce, or your team's CRM) to enrich the dossier, with routes for manual pulls and full syncs. Work the account. Reps set pipeline stage, segment, owner, notes, and manage contacts through small dedicated write endpoints. Outreach drafting. An email-drafting assistant generates and revises outreach copy, split into a broad "strategy" LLM call and a narrower "writer" call so revisions don't resend full research context each time. Fast read models. Maintain a periodically-rebuilt index/cache so list and dashboard views don't scan the full dataset live. Outreach workflows ("flows"). Let reps build ordered outreach sequences (email/call/social/task steps) — generated from a plain-English request via the LLM, or built manually — enroll contacts, and track progress step by step, logging each touch. Lead queue. Pull open leads from the CRM, blend in any intent/engagement signals you have available, score/prioritize them, and surface them alongside the researched account book — reps choose which leads to enroll into outreach, never auto-enrolled. Prospect drafting. A gamified "draft" flow lets reps pick new prospects into their outreach ranked by fit. Starring & saved lists. Reps can star accounts to track and save reusable contact lists for bulk actions. Team spaces. Reps can form shared team spaces so managers/partners see and edit the same accounts without moving ownership — access control checks ownership, explicit shares, and team membership. Compliance. Maintain a company-wide opt-out/suppression registry; every outbound email path checks it and fails closed. Sync opt-out/bounce signals in from your CRM and any outreach-engagement tool (e.g. Gong) on a schedule, and capture opt-outs automatically from reply-to-unsubscribe emails via an inbound email trigger. Cost tracking. Every LLM call logs token/cost usage to a shared log/volume so you can report spend by step, model, and day on the admin dashboard. Backup/restore. Nightly scheduled backup of all account data with an admin-triggered restore/import route, plus migration export/import routes for moving data between environments. Storage: One shared persistent volume holding per-owner account, contact, activity, settings, and index data, plus a separate small volume for cost events and one for backups. Auth: Reps authenticate via your org's SSO; write/read routes are gated to signed-in reps, with a couple of machine-to-machine internal routes locked down by an internal secret rather than SSO (since server-to-server calls can't carry an SSO session). Build it as many small single-purpose API routes plus one or two React UI pages, all wired together as described above — don't overload single steps with multiple responsibilities. Tools used: Anthropic Claude, Salesforce, Gong, Gmail, 6sense, inbound email trigger