Practical playbooks for automating email, research, code delivery, and support with agents, MCPs, and custom pipelines.
I design workflows where AI is the default interface — not a bolt-on — using skills, MCP servers, and custom pipelines to move from idea to production faster.
Every build starts with task decomposition: what should be automated, what needs human review, and which model or tool fits each step.
Production systems combine RAG, agents, structured outputs, and observability so automations stay reliable under real load.
ApproachClassify intent with a lightweight model, route to MCP-connected CRM or calendar tools, draft replies from approved templates, and queue human review for edge cases.
Fetch threads, labels, and send approved replies.
Expose search, calendar, and CRM actions to the agent safely.
Branch on intent, enforce policies, and log each step.
Ground drafts in past threads, playbooks, and product docs.
ApproachIngest sources into a pipeline, chunk and embed with metadata, run multi-step synthesis agents, and export structured briefs with citations.
Normalize PDFs, URLs, and notes into a searchable corpus.
Improve precision before the LLM sees context.
Separate summarize, compare, and critique steps for quality.
ApproachUse repo-aware agents with skills for linting, test generation, and PR descriptions; keep humans on architecture and merge decisions.
Scoped edits with project conventions baked into prompts.
Open PRs, attach summaries, and link issues automatically.
Regression checks on prompts and tool calls before deploy.
ApproachHybrid retrieval over FAQs and tickets, tool use for lookups, escalation rules, and feedback loops into the knowledge base.
Create drafts, tags, and escalations in the existing stack.
Force category, confidence, and citation fields on every reply.
Track deflection, latency, and retrieval misses weekly.