Practical AI – "Creating a private AI assistant in Thunderbird"
Date: September 23, 2025
Host: Daniel Whitenack (B) and Chris Benson (C)
Guest: Chris Aquino (D), Software Engineer at Thunderbird
Episode Overview
In this episode, the hosts dive deep into the journey of bringing an AI-powered, privacy-respecting assistant to Thunderbird, the long-standing open-source desktop email client. Guest Chris Aquino unpacks the motivations, technical decisions, privacy challenges, and practical learnings from building “Thunderbird Assist”—an experimental AI feature aiming to make managing email productive without compromising user autonomy or data security.
Table of Contents
- Introduction & Background ....................................[00:49]
- How AI Is Being Integrated into Email ...............[09:24]
- Technical Approaches to AI in Thunderbird .......[16:00]
- Privacy, Data Flow, and Model Selection ............[21:46]
- Partnering with Flower Labs .................................[26:28]
- Engineering Experiments & Lessons Learned ......[29:04]
- User Needs & Possible Use Cases ........................[37:52]
- Looking Ahead: The Future of AI in Email ...........[45:23]
- Notable Quotes
- Timestamps of Key Segments
Introduction & Background [00:49–09:24]
- Chris Aquino’s Journey to Thunderbird
- 20+ years in web development, including time at SurveyMonkey.
- Joined Thunderbird to work on experimental projects outside the core desktop application.
- Thunderbird’s Place in the Mozilla Ecosystem
- Thunderbird is a 20+ year old open-source email client, spun off from Mozilla Corporation.
- Maintained by the Thunderbird Project under the Mozilla Foundation umbrella.
- Unique in being fully offline-capable: “your email is your own. It gets downloaded to your computer. So if you lost Internet access, you can still read your email.” — Chris Aquino [07:14]
How AI Is Being Integrated into Email [09:24–16:00]
- Trends in AI-Enhanced Email
- Market features: autocomplete, automatic summarization, reply suggestions (e.g., in Gmail).
- Trade-offs:
- Time savings vs. tone loss and dehumanization.
- Privacy concerns with cloud-based processing.
- Key privacy point: “We’re very privacy-respecting, privacy-preserving because a lot of our users choose to use Thunderbird because of that. They want to manage and own their own email.” — Chris Aquino [12:51]
Technical Approaches to AI in Thunderbird [16:00–21:46]
-
Options Considered
- Local embedding of models in Thunderbird: Strictly local/private but resource intensive and not compatible with all hardware.
- Separate local inference application: More secure but adds user complexity.
- Remote API-based inference: Offloads computation, but raises data privacy concerns.
-
Why “Companion” not “Built-in”
- AI features run in parallel to the Thunderbird desktop roadmap to avoid forcing a major intrusive change.
-
Hardware Constraints
- Many Thunderbird users run on older hardware or Linux; heavy local inference is often not feasible.
Privacy, Data Flow, and Model Selection [21:46–26:28]
-
Remote Inference: Privacy Priorities
- Issues: Where does the data go? Is it used for model training? How can user data be protected?
- Explored API providers, requiring guarantees: “No, we’re not using [your data] for training.”
- Flower Labs provided:
- Explicit privacy guarantees (no data retention, not used for training)
- End-to-end encryption
- Model fine-tuning and tooling for evaluation and prompts
-
Quote: “They [Flower Labs] literally can’t help you debug your prompt because it’s encrypted now… That was the fact that they were so helpful at every step. Except when I sent some bad data… they really built it so they can’t see it.” — Chris Aquino [27:18]
Partnering with Flower Labs [26:28–29:04]
-
Origin of the Partnership
- “It just kind of fell into our laps because …Mozilla [is] an investor in Flower. Mark Surman connected Ryan … with Daniel from Flower, and I just ended up on a Zoom.” — Chris Aquino [27:02]
-
Technical Collaboration
- Flower provided an SDK, helped with encryption setup, model selection, prompt-tuning, and rapid iteration.
Engineering Experiments & Lessons Learned [29:04–36:59]
-
Developing ‘Thunderbird Assist’
- Iterative experiments: From basic email summary/reply features to the ambitious “Daily Brief” (an executive summary of emails).
- Challenges:
- Over-prompting led to poor results for complex tasks (“the garbage that I got back sometimes was epic.” — Chris Aquino [31:50]).
- Realization: Break up tasks for the LLM; don’t demand everything in one prompt.
- Hybrid Approach:
- Local Bayesian classifier assigns importance, offloading work from LLM.
- Formatting handled locally to ensure deterministic results.
-
Developer Learning: “The future of this feature is not to just keep on sequentially prompting the same language model for like, hey, now do this, now do this… Instead, …dedicate some small models to specific tasks and then coordinate them in some sort of much more deterministic way.” — Chris Aquino [34:38]
User Needs & Possible Use Cases [37:52–40:38]
-
Current Audience
- Pilot usage is internal (Thunderbird employees only).
- Three main features in ‘Thunderbird Assist’:
- Individual email summarization—effective for long threads.
- Reply generation—for speedier email responses.
- Daily Brief—an overview feature still in development due to technical constraints.
-
Future User Needs
- Semantic search, integration with calendar/tasks/RSS, and cross-information correlation.
- “If we could correlate between these different pools of information, that could be extremely useful.” — Chris Aquino [39:48]
Looking Ahead: The Future of AI in Email [45:23–51:41]
-
Integration into Thunderbird Pro
- Thunderbird Pro = new suite of services (including encrypted file sending and “Thundermail,” Thunderbird’s own email server).
- Envisions offline semantic search and federated learning to pre-generate useful data for users—always with privacy front and center.
-
Expanding Model Context
- Aspirations to let the AI “help you sift through … notes from the last…decades,” connect calendar/tasks/RSS, and recommend relevant info.
- “I would love an LLM to help me sift through that. …As I’m making a note, it could suggest related documents and ideas that I’ve had in the past.”
-
Desire for Determinism and Modularity
- “I want discrete inputs and outputs. I want language models that are small and dedicated to specific tasks. And then I want reusable, shareable ways of wiring them together.” — Chris Aquino [51:27]
- Critiques generic chat-based UIs for most tasks, advocates for more structured, testable workflows.
Notable Quotes
-
On privacy:
“With Thunderbird it is, it’s free and open source, no ads ever, forever. And your email is your own.” — Chris Aquino [07:14] -
On AI trade-offs:
“You’re gaining time. But what you’re trading are things like tone… That email from your mom doesn’t really sound like your mom.” — Chris Aquino [12:12] -
On building AI responsibly:
“Instead, what we’ve done is we have built it, as I’m going to call it, a companion for right now.…We just kind of want to run our AI experiment parallel to [Thunderbird’s] roadmap.” — Chris Aquino [16:10] -
On Flower Labs’ privacy approach:
“They took care of all of our needs. They moved things around on their own development roadmap and gave us early access to things like end to end encryption…” — Chris Aquino [24:37] -
On required AI architecture:
“It’s kind of like when you have a junior developer…all in one function, right? In this AI world, there’s that need for splitting up…makes things more testable and all of that as well.” — Daniel Whitenack [40:38] -
On the future of user-centric AI:
“Make it possible for people to have more control over their information, help them retain their privacy, but…make those creative connections that only they as a human can do.” — Chris Aquino [49:55]
Key Timestamps
- Thunderbird’s Philosophy & History – [04:07–07:14]
- The Role and Risk of AI in Email – [11:06–13:40]
- Technical Paths Explored for AI Integration – [16:00–19:25]
- Partnering with Flower Labs for Privacy – [24:37–28:13]
- Challenges with LLM Outputs and Over-Prompting – [29:04–32:47]
- Hybrid Model Approach and Lessons Learned – [33:00–36:59]
- User Segmentation & Feature Testing – [37:52–40:38]
- Vision for Modular & Deterministic AI Workflows – [50:07–51:41]
Memorable Moment
“So that was an important lesson, was like, okay, so currently, you really need to be very careful, specific and constrained in what you ask the model for.”
— Chris Aquino [33:50]
Closing Thoughts
The episode emphasizes that practical AI in personal communications is about productive user empowerment, not flashy features. Experiments in Thunderbird highlight the importance of privacy, collaboration with trusted partners like Flower Labs, and a shift toward modular, deterministic AI tools that serve—not subsume—human agency.
For more:
Visit Practical AI FM or connect with the hosts for ongoing discussions on making AI accessible and trustworthy.
