I ship
and I show.

Bloort

An AI platform I took from a solo prototype to a paid product on client sites. It started as a RAG chatbot and is now a sandbox for building and testing AI agents. In the live demo you paste a URL, an agent is built in about a minute, and you can talk to it without signing up.

• Sold it to clients and ran it on their live sites. I handled the sales calls, onboarding and billing myself.
• Multi-tenant and secured for production: per-client Pinecone namespaces, layered prompt-injection defence (an LLM guard with a regex fallback), SSRF and XSS filtering, and email and phone masking before anything is stored.
• A semantic cache reuses retrieved context when a new question is at least 0.88 cosine-similar to an earlier one, which cuts Pinecone lookups and LLM cost. Model calls have retries with backoff, timeouts, per-run cost limits, and a dry-run gate on irreversible actions like placing a call.
• Agents use real tools over the Model Context Protocol, including a calendar server for live availability. Outbound VAPI voice calls and Twilio WhatsApp handle lead callbacks and owner alerts.
• Evals run on versioned golden datasets. A trajectory eval caught the agent inventing an appointment date in a tool call, and the fix took the 6-scenario suite from 5 of 6 to 6 of 6. Retrieval evals ruled out a hybrid retriever that did worse than plain dense search (recall@1 fell from 95% to 70%).

• Next.js
• TypeScript
• LangChain
• OpenAI
• Pinecone
• MCP
• VAPI
• Twilio
• Firecrawl
• Firebase
• Railway