Below is the outline for the talk titled “A Practical AI Approach for the Blind Developer” which I presented to the NFB in Computer Science virtual seminar on December 14, 2025.
Introduction
- Michael Forzano
- Senior Software Engineer at Amazon Worldwide Stores Accessibility
- Co-Founder of RS Games, www.rsgames.org
- NFBCS board member
Talk overview
- Everything in this talk I’ve learned over the past year
- AI has transformed the way I work
- Increased productivity? Maybe some.
- Increased effectiveness? Definitely.
Agenda
- Overview of tools
- Vibe coding
- Overview of agentic AI assistant
- MCP servers
- Daily workflows
Overview of Tools
Command Line
- CLI interface to chat with LLM
- Can code for you (create and modify files, run shell commands)
- Can access tools through MCP servers
- Examples:
- Currently using Kiro CLI on WSL
IDE-based Tools
- IDE extension
- Functionality similar to CLI tools
- Example: Cline - most accessible option
Vibe Coding
- Many ways to do it
- True vibe coding: Build me a blogging platform
- You can plan it out, what tech to use, etc.
- AI will go build the whole thing end-to-end in minutes (including build and test)
More Control
- Implement a class that does xyz
- Add unit tests for this class
- Update function x to also do y before returning
CLI vs. IDE
- Cline has plan/act mode
- AI unable to make any changes in plan mode
- Otherwise, can be unpredictable
- AI will ask you questions to clarify requirements
- Even with CLI can ask it to do this, but have to be explicit, might just start making changes
Cline Overview
- Web-based interface inside VSCode
- Install extension
- Command pallet (Ctrl+Shift+P) -> Show Cline
- Plan/act toggles at bottom, edit box, and messages will show above
- Some a11y issues but overall usable
- Tip: Don’t try to watch it make changes
Vibe Coding Caveats
- Might not meet the bar for production ready code
- AI may misunderstand your requirements and start making incorrect changes (especially if not using plan mode)
- Confident halucination
- Context overflow
Agentic AI Assistant: Overview
- Software engineers don’t typically get an executive assistant
- Now they can (sort of)
What it is
- LLM with access to tools through MCP servers
- Access to files (can read and write from them, search directories)
- Agent JSON configuration
- Context (system prompt)
MCP Servers
- Model Context Protocol - open source standard for LLM to communicate
- Server simply takes input from LLM -> does something -> return response
- MCP library available for all major programming languages. Python or Javascript typically used.
- Examples: Connect to Slack, Outlook, task management software like Jira or Assana.
- Just wraps their APIs (i.e. search tool calls search API)
- Many are Vibe coded
Daily Workflows
- Defined in the context/system prompt (markdown file)
- I say “Good morning” -> checks calendar, sprint tasks assigned to me, code reviews outstanding.
- Tells me what tasks need to be done, what meetings I have, creates plan and writes them to work log
- I say “look at what I’ve worked on today” -> checks Slack messages and emails sent, meeting notes, CRs reviewed and submitted
- -> logs my accomplishments and action items to work log
- I say “tell me if there are any Slack messages I haven’t responded to” -> Tells me if there are any in past 3 days
Other Tasks
- Look through the backlog for the projuct I work on and help me find the task about xyz. Easier than UI.
- Update my sprint tasks
- Summarize a codebase or code review
- Query internal websites (wiki, documentation)
- Find the meeting notes from yesterday’s team sync and tell me the decision we made about xyz
Conclusion
- AI is fundamentally changing how we work
- Developers can be the first to incorporate new use cases
- As blind developers in particular, it gives us an advantage