The Brief
The founder was a VP of Technologies & Business Development at a major local firm with deep relationships across the local market - law firms, insurance companies, car dealerships, content sites, health providers. He had the connections and the market access, but no AI engineering talent.
He didn't just want a chatbot. He wanted a platform that could turn his clients' documents and websites into AI-powered customer support - and turn those conversations into qualified leads.
What We Built
A full B2B AI platform with 22 database models, 45 API routes, and 34,000 lines of TypeScript. Not a wrapper around ChatGPT - a production SaaS with multi-tenancy, billing, and a reseller model built in from day one.
Professional-Grade Content Ingestion
The platform ingests documents, websites, and audio into a searchable knowledge base. PDFs are processed with Gemini-powered OCR - parallel batch processing that handles scanned documents, extracts image descriptions, and preserves page structure. Excel files are parsed sheet-by-sheet with table structure intact. Word documents, CSV files, and raw text each have format-specific processing with semantic chunking.
Beyond file uploads, the platform crawls websites on a schedule - daily, weekly, or monthly - via Google Cloud Scheduler. Knowledge bases stay fresh automatically without manual re-uploads. Each crawl job tracks errors per URL and persists scraper configuration.
Multi-LLM Chat with Hybrid Search
The chat interface supports 10+ models across three providers: Google Gemini (1.5 Pro/Flash, 2.0/2.5 Flash, 2.5 Pro), OpenAI (GPT-4o, GPT-4.1), and Anthropic (Claude 3.7 Sonnet/Haiku). Responses stream in real-time with citation tracking - users see which documents and pages the answer came from.
Retrieval uses a hybrid approach: vector similarity search via pgvector combined with BM25 ranking for precision. Admins can tune the retrieval depth per agent to balance context quality against response speed.
The platform also supports audio: users upload voice recordings (WAV, MP3, M4A, OGG, FLAC) that are transcribed via OpenAI and processed as queries.
AI-Powered Lead Collection
This is what makes the platform a business tool, not just a chat tool. The platform monitors conversations in real-time, classifies when a visitor is showing lead intent, and extracts structured parameters from the conversation - name, email, phone, interest, whatever the client configures.
Leads are submitted to external CRMs via webhooks with full authentication support (OAuth 2.0, Basic Auth, API Key). For a car dealership, this means a website visitor asks about a model, the AI answers from the dealership's inventory docs, and the lead lands in their CRM - automatically, with the full conversation attached.
Embeddable Widget
A script-tag widget that clients drop onto their websites. The widget supports RTL languages (Hebrew, Arabic), light/dark themes, mobile-specific positioning, and Google Tag Manager integration. Conversation digests are emailed to subscribers on a schedule.
Two widget versions (v1, v2) with customizable positioning, consent banners, and introduction text give clients control over the experience without touching code.
Multi-Tenant Platform
Every client gets their own organization with isolated data, agents, and knowledge bases. A granular permission system controls who can add users, manage data, configure agents, and edit settings. Token-level cost analytics break down usage by model and provider, with tiered pricing calculations.
The Result
The platform is now serving 10+ enterprise clients across law, insurance, automotive, and healthcare. It handles real production traffic with real business data - ingesting documents, answering customer questions, and generating leads around the clock.
The founder's market access combined with M87's engineering created a platform that scales through relationships, not marketing spend. Each new client is a new organization, configured and live in hours.