The right one for the job.

WARM uses a multi-model architecture — we route to different AI providers based on the task:

• Anthropic Claude — Primary model for email generation. Best-in-class for natural, professional writing with the lowest hallucination rate. When words matter, Claude writes them.

• OpenAI GPT — Used for specific parsing, classification, and structured data tasks where it excels.

• Locally-built models — Internal operations like routing logic, intent detection, and preprocessing. Faster, cheaper, no external API dependency for non-critical paths.

Why not just pick one? Because no single model is best at everything. GPT-4 is great at some things, Claude is better at others, and local models are faster and cheaper for tasks that don't need the heavy artillery.

We're model-agnostic by design. When better models drop, we plug them in. You get the best available tech without waiting for us to rebuild everything.

Asynchronous batch processing. When you trigger a SCRUB scan, we don't lock up your browser waiting for 50,000 contacts to process.

Here's how it works:

1. You click "Run Scan" — job gets queued

2. SCRUB processes records in parallel batches (chunked to avoid timeouts)

3. Progress updates in real-time on your dashboard

4. Results populate as they complete — you can start reviewing before it finishes

5. Final report delivered when complete, also emailed to you

For large datasets (100k+ records), the full scan might take 10-15 minutes, a million records might take several hours. But you're not waiting — it runs in the background while you do other things. Most CRMs under 50k contacts complete in under 5 minutes.

This async architecture also means we can scale. Whether you have 500 contacts or 500,000, the system handles it without choking.

WARM isn't one AI — it's a system of specialized brains, each optimized for a specific job:

• Opener Brain — Researches prospects, crafts contextual first touches

• Chaser Brain — Writes intelligent follow-ups, knows when to stop

• Lazarus Brain — Analyzes closed-lost context, crafts re-engagement sequences

• Hunter Brain — Autonomous lead discovery (coming soon)

• Compliance Officer — Validates every output before sending

Each brain has its own system prompt, context window, and decision logic. They hand off to each other based on triggers and outcomes. It's not one prompt doing everything — it's an orchestrated system.

Every AI-generated email passes through our Safety Fork before sending. It's a branching validation layer that:

1. Checks for hallucinated facts (pricing, features, claims)

2. Validates tone against your brand settings

3. Flags low-confidence outputs for human review

4. Blocks anything that fails validation

5. Routes bad outputs to error logging, not your prospect's inbox

If the Compliance Officer catches a problem, the email never sends. Period.

Native OAuth 2.0 for both HubSpot and Salesforce — no API keys, no CSV uploads, no middleware. When you click "Connect," you authenticate directly with your CRM. We request only the scopes we need and handle token refresh automatically.

Data flows both ways:

• Inbound: Contact records, company data, deal stages, notes, activity history

• Outbound: Email sends logged to contact timeline, activity synced back to CRM

Emails appear in your CRM as if your rep sent them manually. Full timeline visibility, no black hole.

We currently support SalesForce and Hub Spot. If you have another CRM you want integration for let us know and we can have it implemented in less than a week.

Real-time for outreach operations. Changes in your CRM reflect immediately in WARM. When a deal closes, a contact opts out, or a note gets added — we see it instantly via webhooks. No overnight batch jobs, no "wait 24 hours."

This matters for Lazarus especially. If a closed-lost deal reopens, we stop the re-engagement sequence immediately. Batch sync competitors might send an awkward "just checking in on that dead deal" email to someone who just signed a contract.

Webhook listeners detect:

• Hard bounces — Contact marked invalid, sequence stops

• Soft bounces — Retry logic with backoff

• Replies — Sequence stops immediately, rep notified

• Unsubscribes — Contact flagged, all sequences stop

• Out-of-office — Detected and handled appropriately

No manual monitoring required. The system responds faster than any human could.

• Frontend: Next.js 14 (App Router)

• Backend: Supabase (PostgreSQL + Edge Functions)

• Workflows: n8n (self-hosted)

• AI: Anthropic Claude API, Open AI API, Custom local

• Hosting: Vercel

• Payments: Stripe

• Email: Customer's own CRM email API (not third-party SMTP)

Everything is serverless, auto-scaling, and built on infrastructure trusted by enterprises.

PostgreSQL Row-Level Security (RLS) at the database level. Every table has policies that filter by tenant_id. It's not application-layer filtering that could be bypassed — it's enforced by the database itself.

Customer A literally cannot query Customer B's data, even if there's a bug in our code. The database won't return it.

Can I customize the AI's tone?

Yes, at multiple levels:

1. Company default — Set the baseline tone (Professional ↔ Friendly slider)

2. Agent defaults — Configure different personalities for different use cases

3. Rep extensions — Chrome/Firefox extensions let individual reps set their own tone, sign-off, bio, and working hours

4. Disqualification criteria — Tell the AI who NOT to pursue (competitors, students, companies under X employees)

Admin settings are the source of truth. Rep settings personalize within those guardrails.

Each brain is optimized for its context needs:

• Opener — Uses extended context to pull in research, CRM history, company profile

• Chaser — Lighter context, focused on previous thread and response signals

• Lazarus — Deep context including original deal notes, close reason, timeline

• Compliance — Full context of the draft plus company facts for validation

We manage token budgets automatically. You don't need to think about it.

This is WARM Building Principle #1: Always assume messy CRM data and lazy reps.

Every workflow has fallback layers:

• No contact name? Fall back to company name

• No close reason? Check the notes field

• Empty notes? Use deal stage + timeline context

• Missing email? Skip and flag for SCRUB

• Comma-separated fields? Parsed correctly, not treated as garbage

The SCRUB engine runs 47 automated checks before outreach:

• Duplicate detection

• Invalid email syntax

• Likely bots (fake names, suspicious patterns)

• Orphaned contacts (no company association)

• Dead emails (previous hard bounces)

• Zombie deals (closed-lost with re-engagement potential)

• Missing required fields

• Data standardization issues

We don't fail silently. We degrade gracefully.

Yes. We pull custom properties from both HubSpot and Salesforce. Map them in your company profile, and the AI can reference them in outreach.

For Salesforce specifically, we're building Metadata API support for auto-provisioning custom fields to match our HubSpot Properties API capability (coming Q1).

Built on Vercel and Supabase — both have 99.9%+ uptime SLAs. We inherit their reliability. No self-managed servers to go down at 2am.

Workflow execution (n8n) is self-hosted for control and cost efficiency, with monitoring and auto-restart configured.

Not yet for external use. The platform is API-first internally, so a public API is on the roadmap once we stabilize post-launch. If you have a specific integration need, talk to us.

Yes. Full data export available anytime. Your data is yours — we don't hold it hostage.