Curated AI Work

Why This Matters

Built a robust automation backbone for a scaling fintech startup. We combined the visual observability of n8n with the stateful logic of LangGraph to create agents that can handle multi-day research tasks, human-in-the-loop approvals, and automated error recovery.

Implementation Notes

• Designed a 'Self-Healing' node pattern in n8n that automatically retries failed API calls with exponential backoff.
• Implemented LangGraph 'Time Travel' allowing admins to roll back agent states and debug logic mid-execution.
• Reduced operational overhead by 80% by automating the entire customer onboarding and KYC verification flow.
• Custom MCP tools allow the agent to interact directly with internal DBs and Slack without exposing credentials.
• Deployed on a private VPS using Docker, ensuring 100% data privacy and SOC2 compliance readiness.

Why This Matters

A fully production-grade RAG platform rebuilt for enterprise scale. We optimized the retrieval pipeline using hybrid search (Dense + BM25) and cross-encoder reranking, ensuring that the model always gets the most relevant context.

Implementation Notes

• Scaled to 10M+ embeddings with sub-200ms retrieval using LanceDB and HNSW indexing.
• Hybrid Search (Keyword + Semantic) increased retrieval accuracy by 25% over standard vector search.
• Implemented workspace-level RBAC, ensuring strict data isolation for enterprise departments.
• Built a custom 'Source Attribution' UI, allowing users to verify every AI response against original documents.
• Reduced internal knowledge discovery time by 30% across a 500-person engineering team.

Why This Matters

Developed a custom outbound agent that mimics the workflow of a top-tier SDR. It researches companies using live web search, identifies key pain points, and crafts hyper-personalized emails that don't look like they were written by AI.

Implementation Notes

• Tavily Search integration allows the agent to reference news and events from the last 24 hours.
• Sophisticated prompt-chaining ensures emails focus on 'Business Value' rather than just 'Product Features'.
• Human-in-the-loop dashboard allows sales leads to approve or edit drafts before they are sent.
• Direct integration with CRM systems (HubSpot/Salesforce) for seamless lead management.
• Resulted in a 4x increase in meeting book rates and a 60% reduction in lead research time.

Why This Matters

A production-grade AI support chatbot that uses a 'Confidence Threshold' to ensure accuracy. If the model is unsure, it proactively escalates to a human agent in Slack with a summary of the issue.

Implementation Notes

• Confidence Scorer ensures only high-accuracy answers reach the customer, maintaining a 4.8/5 CSAT.
• Automated Slack Escalation routes complex issues to the right department with full context summaries.
• Self-learning loop: Human corrections are automatically fed back into the knowledge base for future queries.
• Native integration with Zendesk and Intercom for a unified support experience.
• Successfully deflected 60% of tier-1 support tickets, saving approximately $15k/month in support costs.