TL;DR

We built an AI study assistant that:

• Watched Firestore for new chat docs
• Spun up multiprocess "agents" to handle each request
• Mixed GPT-4 (8k context) with live web search and file uploads
• Let the model plan its own next steps, then let the user approve or edit

We hit 1,200 users and about $500 MRR before pausing the project.
Here's what worked, what didn't, and where the opportunity still sits.

1. How the engine worked

The system followed a multi-step process:

1. User input: User enters a question or uploads a PDF
2. Document creation: Creates a Firestore document with "pending" status
3. Process activation: A listener detects the document and spawns a dedicated process
4. Planning phase: The system parses the input and creates a step-by-step plan
5. Execution phase: Performs research, coding, writing, or calculations as needed
6. Status updates: Updates the document as each step completes
7. Continuation or completion: Either returns to planning for more steps or marks as complete

Highlights

• Agentic loop: every step produced a fresh instruction list; users could pick the next move or inject their own.
• File and URL support: PDF summarizer + Google Custom Search + relevance filter. We were first to market with PDF uploads in a chat interface, months before ChatGPT added this feature.
• Cheap workers: GPT-3.5 for heavy lifting, GPT-4 only when precision mattered.
• Process fences: each doc got its own Python process with a stop_event so "stop" and "resume" felt instant.

Pain points

1. Context cost: an 8k GPT-4 call was around $0.90. One multi-step session could torch several dollars.
2. Over-engineering: dependency graphs, live logging to Firestore, AutoGPT-style recursion... fun, but 80% of users just wanted fast summaries.
3. Ops overhead: scraping, Chrome-less requests, DALLE images, embedding batches: lots of moving parts, lots of 5× cost spikes when something looped.

2. Metrics that mattered

We could have doubled revenue by halving cost, but the complexity curve was steeper than the growth curve.

3. Entrepreneurial lessons

1. Ship the core loop first Students wanted "upload PDF -> get outline". We delivered that in week one but spent month two on fancy agent orchestration that only power users touched.

2. Watch unit economics early Token fees scale linearly; your price tier probably doesn't. Lock the LTV / COGS ratio before you fall in love with a feature.

3. Kill features fast DALLE illustrations were fun demos, but less than 3% of chats used them. We should have axed the call entirely instead of patching image-handler bugs.

4. Cheap > smart (for now) Even a less capable model becomes magical when the per-token price drops by 10×. A future sub-cent GPT means:

   • Whole-textbook embeddings
   • Real-time code tutors
   • Personal "AutoGPT" agents that iterate all night for pennies

4. What I'd do next time

5. Final word

Studyhub.ai paused, not failed. The market is still wide open and GPT prices keep falling. When a ten-page GPT-4 answer costs less than a penny, someone will nail the friction-free study coach.

One of the toughest realizations was the lack of defensibility in the AI app space. We pioneered file uploads and document Q&A months before ChatGPT released the same feature. When they launched it overnight, our key differentiator vanished. This pattern repeated across the ecosystem: build something innovative today, see it as a free feature in ChatGPT tomorrow.

The final blow came when OpenAI released custom GPTs, allowing anyone to create specialized assistants with minimal technical knowledge. Our sophisticated agent infrastructure, which took months to build, could now be approximated by a GPT with a simple custom prompt and a few PDF uploads. What had been our technical moat became a configuration option in a user-friendly interface.

Next time I'll resist the temptation to build everything because we can and laser in on what users actually pay for, with a focus on proprietary data, workflows, or integrations that create more durable moats.

Keep it lean, watch the meter, and ship the magic.