PipeLance Architecture — Infrastructure to Build and Close Pipeline

Scale

The infrastructure depth enterprise teams expect

500K+ TypeScript LOC across the app source. 200+ named migrations. 300+ database tables and 800+ audit log call sites. Multi-tenancy enforced at the database layer. Security and compliance built into the foundation, not bolted on after the fact.

Codebase

500,000+ lines of production TypeScript across the app source, with 1,827 TypeScript files in the core src tree typed end-to-end.

Strict TypeScript throughout — compiler enforces correctness. No runtime surprises that strict mode would have caught.
Zod at every boundary — API inputs, AI tool definitions, form submissions. Runtime validation everywhere data enters the system.
1,827 TypeScript source files — clean module boundaries across CRM, pipeline assistant, automation, reporting, support, and mobile/API surfaces.
200+ named migrations — every schema change is explicit, versioned, and reproducible from scratch. The question "how do you handle schema changes in production?" has a complete answer.

Why it matters: Codebase scale signals architectural commitment — the schema coverage, domain modeling, and test surface that enterprise evaluators expect before they sign.

API Surface

800+ API route handlers delivering the API depth of an enterprise CRM platform — implemented in a single Next.js application with no microservice coordination overhead.

800+ API route handlers operating under Supabase RLS — multi-tenancy is structural, enforced at the database row level, not an afterthought or a reminder in code review.
RLS enforced at Postgres level — even if application auth fails, cross-org data leaks are blocked by the database itself.
800+ audit log call sites with logAudit() — records user, org, resource, IP address, and operation details on state-changing workflows.
Type-safe end-to-end — database types generated from schema, API types shared, UI types inferred. Breaking changes caught at compile time.

Why it matters: Most CRMs bolt on multi-tenancy. Here it's enforced at the database row level — impossible to query across organizations.

Pipeline Optimization Layer

100+ tools that turn pipeline intent into action

Natural-language requests become governed actions across accounts, campaigns, deals, proposals, calls, and follow-up. Every tool is typed, Zod-validated, permission-scoped, sandboxed for demo mode, and audit-logged through the orchestrator.

Multi-Provider Orchestration

Tiered Claude 4.x routing with GPT-4.1 fallback and exponential backoff. Model-agnostic by architecture — switching providers requires no application code changes.

Unified provider abstraction — new models can be adopted without touching orchestration logic.
Automatic failover — provider errors trigger fallback within the same request. Users see no interruption.
Streaming responses — token-by-token streaming to UI. ReadableStream pipes directly to React state.
Context injection — dynamic system prompts include user role, org context, current page, and relevant entity data.
Structured outputs — Zod schemas define expected response shapes. Invalid outputs retry automatically with schema context.

Acquirer note: LLM costs are a line item on every AI company's P&L. Provider abstraction means costs can be optimized without a re-architecture.

100+ Pipeline Action Tools

The platform doesn't just answer questions - it executes. Tools move pipeline stages, send contracts, book meetings, calculate commissions, log calls, and pull territory performance.

Pipeline & CRM — search_deals, update_stage, create_contact, analyze_pipeline, forecast_revenue, territory_assign
Execution — send_email, send_contract, book_meeting, enroll_sequence, create_deal_room, generate_proposal
Intelligence — summarize_call, extract_action_items, score_lead, find_similar_deals, pull_intent_signals
Finance — calculate_commission, run_clawback, export_payroll, generate_forecast_report

Why it matters: Adding new capabilities means defining a Zod schema and handler. The extension surface is wide open.

Action Authority Mode — Configurable Confirmation Gates

NonePipeLance executes immediately. For low-risk, reversible actions like logging notes or updating deal stages.

ImplicitPipeLance shows what it will do, executes if no response in 5 seconds. Middle ground for standard operations.

ExplicitRequired confirmation before execution. Mandatory for send_email, delete_*, bulk operations, and contract sends.

RollbackReversible operations store full rollback data. 60-second undo toast appears after execution.

Demo ModeEvery tool is sandboxed — full action capability, zero data mutation. Safe for prospects and evaluators.

Audit LogEvery tool execution logged with user, org, IP, input payload, output payload, and tool chain path.

Security & Multi-Tenancy

Defense in depth. Enforced at the database.

RLS policies mean a bug in the application layer cannot leak data across organizations. Security here is structural, not behavioral.

Multi-Tenant Architecture

Hundreds of route handlers operating under Supabase RLS. Cross-org data leaks are blocked at Postgres level, not just in application code.

Row Level Security everywhere — not a single table without org policy. Impossible to cross-tenant query, regardless of application bugs.
7-level RBAC hierarchy — owner → admin → manager → support → rep → member → external. Visibility rules cascade properly.
Connection pooling — Supabase Pooler prevents connection exhaustion. Traffic spikes handled without incident.
Demo schema isolation — separate demo schema mirrors production structure. Sandboxed trials without touching production data.

Why it matters: Enterprise buyers ask for architecture diagrams. This one holds up because the security model is the data model.

Compliance Architecture

800+ audit log call sites record user, org, resource, IP, and operation details across state-changing workflows. GDPR right-to-deletion implemented.

Commission audit trail — append-only log with dispute resolution. Every payout traceable to source deal, rep, and rule.
GDPR deletion endpoint — PII-scrub that preserves record shell for audit FK integrity. "Right to be forgotten" without breaking historical data.
Webhook security — HMAC signature validation on all inbound webhooks (Stripe, Twilio, Zoom). Replay attack prevention with idempotency keys.

These are the features that get a security questionnaire signed. Most CRMs don't have them at all.

Guided Execution

Natural language → multi-step pipeline actions

Intent classification, tool selection, execution planning, and rollback for pipeline workflows. With human approval gates for high-impact actions.

Intent to Execution

"Research Acme Corp, draft an intro email from our last meeting notes, and schedule a follow-up for Thursday" becomes a four-tool chain executed from one command.

Intent classification — user input parsed into structured intent with entity extraction. Ambiguity triggers clarifying questions, not failures.
Tool chaining — model selects and sequences tools. Complex requests become explicit execution plans shown to the user before running.
Parameter inference — missing parameters inferred from context (current deal, logged-in user, org settings). Minimal user friction.
Approval gates — high-impact tools require explicit confirmation. Users see what will happen before it does.

Agent-Assisted Development Infrastructure

PipeLance is built and maintained with agent-assisted continuous improvement loops - the same guided execution architecture the platform delivers to customers, applied to its own development lifecycle.

Continuous quality loops — an audit agent maps gaps against an enterprise checklist, domain agents execute targeted improvements, conflicts resolve cleanly. The codebase improves on every cycle.
Self-extending architecture — new capabilities, integrations, and verticals ship without disrupting the core data model. The architecture scales with the business.
Domain model validated at scale — 200+ migrations, 800+ API route handlers, and 35 platform entitlements have stress-tested the schema. There is no greenfield risk here.
Integration test suite — regression coverage on the highest-risk API routes: auth, contacts, deals, sequences, signatures, commissions, and the pipeline assistant.
35 platform entitlements in code — CRM, AI, email sync, scheduling, call intelligence, deal rooms, enrichment, ICP, automation, reporting, and pipeline execution are all represented in the current entitlement surface.

Pipeline Engine

The capabilities leaders need to create, build, and close pipeline

Forecasting, sequence automation, call intelligence, territory management, and commission tracking unified in a single platform.

Forecasting (Clari-Level)

Pulls deal health scores, stage transition probabilities, historical rep accuracy over 8 periods, and pipeline movement waterfall charts.

Commit / best-case / weighted views — all three computed with the same methodology Clari charges $200/seat for.
Manager alerts — triggers when commit changes >20%. No more spreadsheet reconciliation at end of quarter.
Period-over-period comparisons — 8 historical periods for rep accuracy calibration. The model learns each rep's sandbagging pattern.
Pipeline waterfall — tracks deals added, advanced, slipped, and closed within any period. CFOs can explain variance in real time.

Commission & Territory

Closes the loop between CRM data and finance — most CRMs stop at "track deals." PipeLance goes all the way to payroll export.

Split commissions — fan out to all deal reps. Territory rules auto-assign contacts and fire conflict alerts on multi-match.
Clawback rules — automatically reverse commissions when a won deal is lost within configured window.
Payroll export — runs on schedule. Commission data flows to finance without manual reconciliation.
Append-only audit trail — every commission calculation is traceable. Dispute resolution has a complete source-of-truth.

Sequence Engine (Outreach-Parity)

A/B tested, timezone-aware, CAN-SPAM compliant sequence automation with statistical significance gates.

Timezone-aware sending windows — per-recipient. Emails arrive during business hours without manual scheduling.
A/B variant selection — chi-square significance gates (p < 0.05, minimum 100 sends per variant). No winner declared on small samples.
Per-org daily send caps — audit-log-driven counting. Protects deliverability at the org level, not just the rep level.
CAN-SPAM compliance — unsubscribe injection globally enforced across all sequences in org. Not opt-in — required.

Call Intelligence Pipeline

Provider-agnostic transcript ingestion. GPT-4 analysis extracts structured insights — not just summaries.

Multi-source ingestion — webhook endpoints accept transcripts from Zoom, Gong, Dialpad, or any provider. Normalized to consistent format.
Async processing — analysis runs in background. UI updates via real-time subscription when complete.
Coaching metrics — talk ratio, question frequency, monologue length, engagement score. Managers see rep performance across all calls.
Auto-task creation — commitments made in calls become tasks automatically. Nothing falls through the cracks.

Real-Time & Integrations

278 Realtime channels. 70+ integration modules.

Instant updates across users and devices. Bidirectional integrations — not read-only connectors.

Real-Time Architecture

278 Supabase Realtime channel references across the frontend. Live pipeline updates, notifications, presence indicators, and collaborative record editing — via WebSocket, not polling.

Table subscriptions — subscribe to INSERT, UPDATE, DELETE on any table. Filtered by RLS automatically — users only receive events they're authorized for.
Presence indicators — live indicators showing who's viewing a deal, editing a contact, on a call.
Optimistic updates — UI updates instantly on action. Server confirms or rolls back.
Cross-tab sync — BroadcastChannel API syncs state across browser tabs. Edit in one, see in all.

Integration Layer

70+ modules in src/lib/ covering the full revenue ecosystem. Six inbound webhook handlers with HMAC verification, replay protection, and idempotency.

Bidirectional pipelines — not read-only connectors. Slack, Twilio, Resend, Stripe, Google Calendar, Gmail, Zoom, HubSpot, Salesforce, Pipedrive, Zapier.
HMAC webhook security — signature validation on Stripe, Twilio, Resend, Zoom, HousecallPro. Replay attack prevention with idempotency handling.
OAuth token management — centralized refresh with 5-minute expiry buffer. No expired token errors in production.
HubSpot & Salesforce import — bidirectional data sync. Full migration paths for incoming customers.

Application & Deployment

One unified platform. Ships day one.

No webhook plumbing between seven vendors. No data sync lag. No per-seat license stacking. One production-grade Next.js application with the combined depth of HubSpot, Gong, Outreach, Clari, Zendesk, and Calendly — without the fragmentation.

Application Layer

React Server Components reduce client JavaScript by 60–70%. App Router provides nested layouts, streaming, and parallel data fetching.

Next.js 16.2.1 App Router — Server Components as default, client components explicit. Partial prerendering for instant navigation.
React 19.2.3 — concurrent rendering, useOptimistic for instant UI, useActionState for form handling. Strict mode enabled.
206 accessible components — built to WCAG AA. Focus rings on every interactive element, typed skeleton loaders across all major views (contacts, onboarding, messages, settings, reports, activity timeline), form error states, WCAG AA contrast validated across light and dark themes, mobile drawer sidebar. Ships accessible on day one — most CRMs charge enterprise customers extra for this.
Cursor pagination everywhere — no OFFSET/LIMIT performance degradation. Scales to millions of records without slowing down.

Clean IP & Portability

Unencumbered IP. Fully portable infrastructure. No vendor lock-in on any layer of the stack.

Clean IP — 100% proprietary application code with an unencumbered license profile. Full source available for technical review.
Dependency profile — 98%+ MIT/ISC/Apache-2.0. No GPL in the application layer. Dependency tree fully auditable.
Infrastructure portability — Supabase migrates to self-hosted Postgres. Vercel moves to any Node.js host. LLM providers are abstracted and swappable.
CI/CD pipeline — TypeScript strict mode + ESLint on every push. Preview deployments for every PR. Instant rollback on any deployment.

Why it matters: Portable infrastructure and clean IP remove the integration risk that slows enterprise procurement. The stack can be evaluated, audited, and deployed without dependency on any single vendor.

The Optimization Advantage

PipeLance uses the same guided execution architecture it sells to operate its own development process. Continuous improvement loops map gaps against an enterprise checklist, execute targeted fixes, and verify quality - the same pattern customers use to optimize pipeline workflows.

The result is a platform where the architecture, the domain model, and the security posture have all been stress-tested at scale before a single customer signs. Enterprise infrastructure does not require years and eight-figure budgets when the development process itself is optimized.

Intent to action. At every layer of the stack - from the pipeline assistant a rep uses to the continuous improvement loop that maintains the codebase behind it.

Platform Portability

Designed for flexibility. Supabase migrates to self-hosted Postgres. Vercel moves to any Node.js host. LLM providers abstracted. API-first architecture supports embedding and white-label deployments. Fully functional sandbox environment available for evaluation. No vendor lock-in on any layer of the infrastructure.

Infrastructureto Build and Close Pipeline