Sahai
Enterprise-grade AI support platform that scales customer service
Project Details
Support teams spend 60% of their time answering the same repetitive questions. Meanwhile, AI chatbots frustrate customers with scripted, generic responses that feel nothing like real help. Businesses were stuck — expensive human teams or broken bots. There was no intelligent middle ground.
Sahai gives businesses AI agents that actually understand their documentation, handle conversations naturally, and know exactly when to bring in a human. The result: routine questions get answered instantly, and your team only steps in when it genuinely matters.
What It Does
RAG-Powered Knowledge
1536-dimension embeddings search your docs so the AI answers from your actual content — not hallucinations.
Smart Escalation
Detects frustration, unanswerable questions, and explicit requests — then routes to a human before things go wrong.
Real-Time Dashboard
Live conversation feed with full customer context: location, device, referrer, and AI-assisted reply enhancement.
Voice + Chat
One unified platform for text and voice support. Phone numbers and web-based voice calls via Vapi, side-by-side with chat.
One-Line Embed
A single script tag adds Sahai to any website — React, WordPress, or vanilla HTML. No framework conflicts, no dependencies.
Enterprise Security
Multi-tenant isolation with AWS Secrets Manager BYOK. Every tenant's credentials encrypted, scoped, and never exposed.
In Action
Highlights
What We Built
Sahai is a production-grade B2B AI support platform built in 8 weeks as a Turborepo monorepo — three apps sharing one codebase: an operator dashboard, a customer-facing widget, and a lightweight embed script that drops into any site with a single line of HTML. The AI is powered by OpenAI GPT-4o-mini with a full RAG pipeline (vector search over org-scoped embeddings), real-time message sync via Convex, and voice support through Vapi.
The Hardest Problems
Multi-tenancy was harder than anticipated — every query, every file upload, every credential had to be scoped by organizationId with zero cross-tenant leakage. Secure credential storage for the "Bring Your Own Key" Vapi model meant routing everything through AWS Secrets Manager. The embed widget needed to run isolated in an iframe communicating via postMessage to prevent CSS conflicts across every possible host site. Getting all three to work together reliably — and pass enterprise security requirements — consumed 20% of the entire build timeline.
What We Learned
RAG is 80% prompt engineering. The embeddings model and vector search are commodity infrastructure — the intelligence lives in how you instruct the AI to interpret search results, admit ignorance, and decide when to escalate. We ran 47 prompt iterations on the support agent alone. The breakthrough was stopping trying to make the AI sound "professional" and instead making it genuinely useful: searching when it should, escalating when it can't help, and never fabricating answers.
The Result
Sahai is live and production-ready. Businesses deploy it in under 10 minutes. The platform combines intelligent AI with real-time operator tools — routine questions are handled instantly while the team focuses on the complex, nuanced conversations where humans actually excel. Sahai means "companion" in Hindi. That's the philosophy: AI and humans working together, each doing what they do best.
Like what you see?
Let's build your next product together.