www.bloort.ai/ · demo
01AI agent & interview platform · Founder & Lead Engineer · 2025–present
Bloort
An AI platform I took from a solo prototype to a paid product on live client sites, then kept building. It's grown from a RAG chatbot into an agent sandbox for building and testing AI agents, plus an AI interview platform. The live demo still shows the core loop: paste a URL, watch an agent get built in about a minute, then talk to it with no signup.
- Ran it as a paid product on live client sites. I owned the sales calls, onboarding, and billing, the whole loop from first conversation to a deployed agent.
- Made it multi-tenant and hardened for production: per-client Pinecone namespaces, layered prompt-injection defense (an LLM guard with a deterministic regex fallback), SSRF and XSS filtering, and email and phone PII masking before anything is stored.
- Added a semantic cache that reuses retrieved context when a new question is at least 0.88 cosine-similar to a past one, cutting Pinecone lookups and LLM cost, then wrapped model calls in retry-with-backoff, timeout races, per-run cost ceilings, and dry-run gating on irreversible actions like placing a call.
- Gave the agent real tools over the Model Context Protocol, including a calendar server it queries for live availability, plus outbound VAPI voice and Twilio WhatsApp for lead callback and owner alerts.
- Backed it with versioned golden datasets and a failure-mode taxonomy: an agent-trajectory eval caught the bot hallucinating a tool argument (an invented appointment date) and took the 6-scenario suite from 5 of 6 to 6 of 6, and my retrieval evals rejected a fancier hybrid retriever that scored worse than plain dense search on an already-clean corpus (recall@1 fell from 95% to 70%).
- Next.js
- TypeScript
- LangChain
- OpenAI
- Pinecone
- MCP
- VAPI
- Twilio
- Firecrawl
- Firebase
- Railway