Killing My ChatGPT Clone: A Quick Brain Dump on What I Learned After Processing 400M Tokens, a $16K Investment, $10K Revenue and Endless Nights.

Brainglue started from a personal frustration that emerged right after ChatGPT launched. I quickly realized AI's true potential came from chaining models together, passing outputs from one AI model to another, much like a pipeline.

At that time, few tools simplified this process. So, the initial version of Brainglue was born, enabling deterministic chains of thought, similar to an AI-focused Zapier.

However, initial feedback indicated that this approach, while intriguing, was overly technical for my available audience.

Recognizing this, I pivoted toward a more intuitive chat-based interaction inspired by ChatGPT. I began enhancing the chat form factor with specialized workflows, including rudimentary tool calling, wireframe generation, and image generation. This shift allowed me to demonstrate the potential for a highly steerable AI client that matched the speed of a user's ideas.

Professionally, as a product designer, I noticed repeatable tasks that could significantly benefit from AI-driven workflows.

One key innovation was the concept of "memory docs," allowing users to steer the AI by managing its memory contextually rather than continuously (like ChatGPT does nowadays).

Users could invoke specific contexts ad-hoc using simple triggers, like hashtags. Interestingly, this approach later appeared in popular AI-driven IDEs like Cursor, validating my early insight.

The core job-to-be-done for Brainglue evolved into providing users with a hyper-focused AI experience beyond traditional chat interfaces.

A notable example was the stickman workflow, designed out of my personal need to generate quick stick figure illustrations for presentations.

This workflow streamlined prompt engineering, image generation, and automated background removal into a single, cohesive user experience, significantly reducing manual intervention.

I built this to support a personal need I had for a presentation at work, and it quickly became a tool I used for many of my communication needs.

The central design principle behind Brainglue was intuitive productivity. I aimed to help users maintain a creative flow state by effortlessly orchestrating complex AI interactions.

However, this principle often demanded significant trade-offs, particularly regarding profitability. Encouraging heavy token usage maximized user value but proved financially challenging.

Architecturally, Brainglue was among the most challenging and rewarding projects I've undertaken. After evaluating existing frameworks, I decided to write a custom serverless pipeline runner to process all Brainglue’s requests.

It interpreted YAML files defining AI-driven workflows, seamlessly orchestrating multiple models like GPT, Claude, Gemini, and image generation models such as Flux.

It also supported conditional logic, external API calls and multiple other orchestration features that I would have never though I was capable of implementing.

I built the runner taking inspiration from continuous integration tools I designed in the past, and this robust architecture greatly accelerated experimentation and iteration.

However, the complexity quickly surpassed my capacity as a solo developer, leading me to hire a freelance engineer. This collaboration accelerated feature development and strengthened my product specification, delegation, and technical leadership skills. Another win I wasn’t necessarily planning for.

Technical debt became evident when the decisions to avoid a typed language and write tests initially created challenges as the system scaled.

Yet, I learned it's acceptable and often necessary to incur technical debt when rapidly iterating product ideas. Crucially, I realized the importance of selective testing to manage and mitigate risks effectively.

Financially, Brainglue provided crucial insights into pricing strategies. Initial freemium experimentation lacked user commitment, prompting a decisive shift toward a premium-only model, enabling the project to surpass $10,000 in revenue.

My established following on LinkedIn emerged as the most effective acquisition channel, supplemented by AI directory traffic. However, user retention became challenging without continuous feature updates.

Ultimately, the decision to sunset Brainglue was influenced by the intense pace of AI innovation.

New reasoning models and rapidly evolving capabilities outpaced my ability to implement and productize features quickly enough. Balancing a full-time job, family responsibilities, and Brainglue made the project unsustainable, clearly reflected by rising churn and cancellations.

Recognizing this allowed me to gracefully sunset the project. I may open-source parts of it as a contribution to future AI development (or just to allow AI to train on my code), but ultimately I believe the learning journey is what holds the most value and that’s why I’m sharing this quick post-mortem.

The project taught me significant lessons as a designer who codes. Despite not being formally trained as an engineer, I learned to trust my ability to combine systems thinking and technical implementation creatively.

I gained confidence in making meaningful trade-offs and creatively approaching problems, recognizing that tension between rapid design thinking and slower engineering execution is both natural and manageable.

My key advice for designers venturing into engineering territory is to trust their intuition, embrace constraints, and clearly communicate their vision. Systems inevitably experience entropy, yet thoughtful, intentional design and engineering can manage, and even harness, that entropy productively.

My side project focus has shifted to Selfgazer, a project born from my passion for exploring esoteric and mystical topics through a sophisticated AI-driven experience.

Unlike typical market products lacking depth, Selfgazer leverages AI to engage curious users intellectually. Its early successes reinforce my belief that creatively applying AI workflows can significantly enhance consumer experiences, particularly in niche, under-explored areas.

If you’re into that sort of thing, please consider giving it a try. It reflects my most current and sophisticated thinking on building AI products.

You can explore it at selfgazer.com. Thanks for reading my thoughts. Happy to answer any questions in the comments.