Latest writings
Notes on AI agents, Salesforce, automation, micro-SaaS, and engineering.
SalesforceDay 14/15 — Production Architecture: Governance, Monitoring & Error Handling
Take a Salesforce AI assistant to production: who approves tools, what Salesforce logs, how to classify and retry errors, and which limits and costs to plan for.
SalesforceDay 13/15 — Headless 360 + Agentic AI Architecture
One assistant, a planner with tools, or several agents? Where Agentforce fits, what the beta Headless 360 MCP Server changes, and native React versus Next.js.
Day 12/15 — Custom MCP Tools + Apex/Flow Business Logic
Publish your own Apex and Flow logic as tools on a custom Salesforce hosted MCP server, so the model calls your business rules instead of inventing them.
SalesforceDay 11/15 — Security Architecture: OAuth, Permissions & Least Privilege
A threat model for AI clients on Salesforce, the layers that contain them, where tokens go, why there is no integration user, and a least-privilege permission set explained line by line.
SalesforceDay 10/15 — Build Your First AI + Salesforce Workflow
Turn a chat into a workflow: a read-only meeting prep brief built with Next.js, the Claude API's MCP connector and Salesforce hosted MCP servers.
SalesforceDay 9/15 — Create & Update Salesforce Data Safely
Let Claude write to Salesforce without losing control: what still runs on every write, the propose, confirm, execute, verify loop, and a confirm-before-write pattern for your own app.
SalesforceDay 8/15 — Read Salesforce Data Using Natural Language
Ask Salesforce questions in plain language and trust the answers: how Claude picks soqlQuery, find or getRelatedRecords, prompt patterns, verification and limits.
SalesforceDay 7/15 — Explore Salesforce MCP Tools & sobject-all
What Claude can actually call in your org: the 11 sobject-all tools, the smaller servers, how to read a tool definition, MCP Inspector, and the beta Headless 360 server.
SalesforceDay 6/15 — Connect an AI Client to Salesforce via MCP
Connect Claude to Salesforce's hosted MCP servers: a custom connector in claude.ai and Claude Desktop, a direct server in Claude Code, and your first read-only prompts.
SalesforceDay 5/15 — OAuth 2.0, PKCE & External Client App Setup
Create the External Client App every MCP client needs: the mcp_api and refresh_token scopes, PKCE, JWT tokens, and a tested PKCE helper in TypeScript.
SalesforceDay 4/15 — Enable & Configure the Salesforce MCP Server
Switch on Salesforce's hosted MCP servers, part of Salesforce Headless 360 (AIforce): the exact Setup path, which server to start with, the right URL for your org, and a server inventory.
Day 3/15 — MCP Fundamentals: Client, Server, Tools & Resources
The Model Context Protocol behind Salesforce Headless 360 (AIforce): hosts, clients, servers, tools, the stateless 2026-07-28 lifecycle, and a local MCP server you build and inspect.
SalesforceDay 2/15 — Headless 360 Architecture Explained
Follow one request through Salesforce Headless 360 (AIforce): client, identity, MCP server or API, platform and back. Plus when to use MCP and when to call the APIs directly.
SalesforceDay 1/15 — What Is Salesforce Headless 360 and Why Does It Matter?
Salesforce Headless 360, now called AIforce, opens the platform to agents and code. What it is, what the rename changed, what is GA or beta, and how to get ready.
Why Use Gemini for AI Automation: A Builder's Comparison
When Gemini is the right model for automation workflows and agents: current models, verified pricing, native multimodal input, long context, tool calling, and where other models fit better.
Graph Engineering For AI Agents: My Production Playbook
Stop wiring agents with brittle scripts. Learn how I structure knowledge graphs and state machines to build AI agents that actually scale in production.
AI Agents in Enterprise Automation: A Founder's Field Guide
Stop building brittle RPA bots. Here's how to design AI agents that actually survive contact with enterprise systems—without burning your roadmap or budget.
Building Long-Term Memory: Setting Up Vector Databases for Semantic Search in LangChain
LLMs have a context limit; your data doesn't. Learn how to implement vector databases in LangChain to enable semantic search and build robust RAG pipelines.
Unmasking the Black Box: Mastering LangChain Callbacks for Debugging
Master LangChain callbacks to debug AI agents. Learn to build custom handlers, trace execution flows, and fix 'black box' LLM issues.
Best Practices for Deploying LangChain Apps to Production
Serving, streaming, per-user state, caching, tracing, evals, secrets and timeouts: a checklist for moving a LangChain 1.x agent out of the notebook, checked against the current docs.
Beyond Built-ins: Architecting Custom Tools in LangChain
Standard LangChain tools handle the basics, but production agents need to touch internal APIs and proprietary logic. Here is a deep dive into building robust, Pydantic-validated custom tools.
Architecting Autonomous Agents: A Builder’s Guide to Complex Workflows
Moving beyond simple chatbots requires a shift in architecture. Here is how to engineer autonomous agents using ReAct loops, tool usage, and state management to handle complex, multi-step objectives.
Building State: A Deep Dive into LangChain Memory for Conversational AI
LLMs are stateless by default. To build a conversational agent that actually remembers, you need to manage context effectively. This guide covers implementing various memory types in LangChain, from basic buffers to summary-based retention.