AIdeas Finalist - Diverge — AI-Powered Decision Intelligence
Diverge takes a life decision you are stuck on and runs it through a structured AI debate. You type something like "Stay at my job vs. take the offer" or "Stay in Boston vs. move to Austin." The app spawns two agents, each one representing your future self on one side of that choice. They argue it out across five rounds while pulling from Monte Carlo financial simulations, real salary data, and cost-of-living numbers.
AIdeas Finalist: Diverge — AI-Powered Decision Intelligence
App Category: Daily Life Enhancement
My Vision
The first article was personal. I built Diverge because I wasted two years stuck inside a decision that should have taken two weeks. Spreadsheets, advice, pros-and-cons lists rewritten until the paper wore through. None of it helped because none of it could show me what my life would actually feel like on the other side. I kept choosing safety because safety was easier to explain.
That is still the emotional center of Diverge.
But the finalist round forced me to see the project more clearly. Diverge is not just a tool for my own indecision. It is for anyone standing at a real fork: take the offer or stay put, move across the country or stay near family, keep working or take on debt for school, stay in the stable path or bet on the version of life that feels more alive.
Diverge takes that kind of decision and runs it through a structured five-round debate between two future selves. One argues for the safer path. The other argues for the bolder path. They do not just list benefits. They drag the tradeoffs into the open. They argue about regret, identity, risk, money, comfort, ambition, and what each path quietly costs.
The structure matters. Each round pushes further into the future:
- The Ripple: what happens first
- The Ledger: what compounds over the next few years
- The Mirror: who you are becoming
- The Ghost: what regret and gratitude look like much later
- The Knot: the final synthesis
Nine templates cover the most common forks: career change, new city, startup, education, relationship, lifestyle change, giving back, trade vs. degree, and family crossroads. Users can also build from scratch. Underneath the debate, Diverge mixes emotional reasoning with quantitative reality checks: salary comparisons, cost-of-living data, runway calculations, and Monte Carlo simulations. Between rounds, the user can inject new context like "My partner said they would relocate with me" and the agents fold it into the next response.
Before the verdict appears, a gut check asks the user which way they are already leaning and what they are most afraid of. That answer stays on their device. After the verdict, the app shows it back: "You came in leaning toward X." The point is to make confirmation bias visible. If the AI just told you what you already believed, you should know that.
The verdict itself lays out where each side wins, identifies the blind spot, maps both timelines forward on an interactive fork with nodes at Year 1, 3, 5, and 10, and gives one concrete next move. Curated resources, a book, a video, and a mental model, follow the verdict so the user walks away with something to act on. Email check-ins at Day 7, 30, and 90 bring the decision back when the initial emotion has faded.
That was the product I wanted to build. The finalist process helped me understand the product I was actually building: not an AI that gives advice, but a tool for people trying to think clearly when the stakes are real.
Why This Matters
One of the strongest pieces of judge feedback I received was that the original framing felt too narrow, too close to personal productivity, and too centered on "decision paralysis." They were right.
The phrase "daily life enhancement" can sound small until you think about what actually shapes a life. A promotion accepted or declined. A city left behind or stayed in. A degree pursued at the cost of debt. A caregiving choice that changes the next five years. A layoff that forces someone to decide whether to reskill, relocate, or start over.
Those are not productivity problems. They are mobility problems, opportunity problems, family problems, identity problems. They are the kinds of choices that decide who gets access to what future.
Not everyone has a mentor. Not everyone has a therapist, a coach, a financially literate parent, or a network of people who can help them think clearly under pressure. Plenty of people have internet access and a thousand scattered pieces of information. Far fewer have a process that can hold emotional reality and practical tradeoffs at the same time.
That is the gap Diverge is trying to close.
I looked hard at what already exists. ChatGPT and Claude are powerful but generic. DecisionMe and Rationale are stronger on criteria and bias reduction. Reflection is strong on self-understanding. Levels.fyi and ProjectionLab are excellent at one slice of reality. None of them combine emotional reasoning, financial data, structured debate, and a saved decision record in one flow for ordinary people facing high-stakes life choices.
That distinction changed how I think about impact. Better decisions affect where people live, what careers they pursue, whether they take educational risks, and whether they act on opportunities they would otherwise postpone for years. That is a larger claim than "this helps indecisive people." It is also the more accurate one.
How I Built This
The technical architecture of Diverge reflects the same instinct as the product itself: keep it structured, resilient, and honest about what it actually knows.
The app runs as a serverless AWS architecture with a split backend. The long-running debate workload is separated from the lighter REST operations because a five-round streamed debate can run much longer than a normal API timeout allows. The debate function handles orchestration, token streaming, agent prompts, metrics extraction, and checkpointed progression. The REST side handles templates, journal operations, sharing, and everything else.
On the frontend, React, TypeScript, Vite, and Tailwind. Streaming output batched at 80ms so the page feels alive instead of jittery. The visual language stays cinematic: void black, Cinzel headings, amber and blue path accents. Cloudflare Turnstile handles human verification before the expensive debate workload starts.
Under the hood, the system has three parts that matter most.
First, the debate protocol. The app does not send one giant prompt and hope for something poetic. It runs a structured sequence of rounds with distinct purposes. The safer path and the bolder path are assigned deliberately. Each round feeds into the next. The user can interject between rounds so the process does not feel like a sealed black box.
Second, the heart plus math layer. Diverge is strongest when emotional truth and numerical reality sit in the same room. That is why I built tools around financial context, salary estimates, cost-of-living comparisons, runway thinking, and modeled downside rather than leaving the agents to improvise everything. The goal is not to bury a life choice under fake precision. It is to keep the emotional argument anchored to material facts.
Third, the decision record. A good decision tool should not vanish when the conversation ends. Diverge saves the outcome to a journal, generates a shareable link, and exports a branded PDF. Email check-ins at Day 7, 30, and 90 ask the user to revisit what they decided and compare expectations against reality. That longer learning loop is one of the clearest ways Diverge differs from a one-off AI chat.
The finalist feedback pushed me to think about boundaries. A product like this could drift toward false authority. So the design principle I keep returning to is: Diverge should help users structure hard choices, not surrender them. The gut check before the verdict exists for this reason. It records your intuition before the AI speaks, then shows it back afterward so you can see whether the verdict confirmed your bias or genuinely shifted your thinking. The ability to share a result with a partner, mentor, or counselor is not a convenience feature. It is part of making the product responsible.
The judges flagged my original submission for lacking user validation. Fair criticism. Since then I ran Diverge through real users and an adversarial test suite built to break every input surface: garbage input, prompt injection, XSS payloads, Unicode edge cases, boundary-length fields, category mismatches, and crisis content. Eighty-three edge cases across ten categories, plus full end-to-end debates across all nine templates and custom freeform.
The testing caught real problems. One user ran a relationship scenario and got generic filler instead of a real narrative timeline. When I traced it, the content guardrails were rejecting words like "partner" and "friend" because the user had only typed short paths like "Tell my friend how I feel." The guardrails treated that as too little context and blocked relationship terms entirely, even though the user picked the relationship template. Fix: lower the sparse-context threshold so realistic input stops triggering the strictest mode.
Another user entered only whitespace as their decision paths. The app accepted it and debated nothing. Pydantic enforced min_length=2, but spaces counted as characters. Fix: a field validator that strips whitespace before the length check.
A third bug came from crisis detection. The regex matched "cut myself" but not "cutting myself." Anyone describing self-harm in present tense slipped through. That mattered more than any feature bug. Fix: expand the patterns to handle verb forms, not just base forms. Small change in code. Significant change in responsibility.
The biggest product lesson is that structure beats novelty. "Two agents debate your future" is a fun pitch line. But what makes Diverge useful is the protocol around that idea: the rounds, the interjections, the financial tools, the saved journal, the synthesis, and the ability to come back later with more perspective than you had when you began.
Demo
The demo centers a real life transition and needs to show:
- the five-round structure
- a user interjecting between rounds and changing the next response
- emotional reasoning plus quantitative reality checks
- the final decision record
That is the clearest way to show that Diverge is more than a chatbot with a dramatic prompt.
What I Learned
The judge feedback exposed where my framing was smaller than the product.
The first version was strongest when it was personal and technical. Both parts were real. But the judges were right that I had not explained the competitive landscape, the ethical considerations, or the broader case for why this matters.
The biggest shift was moving from the language of indecision to the language of life transitions. Important decisions are messy hybrid problems — emotional, financial, social, practical, identity-forming all at once. Most tools handle one dimension well. Very few hold them together.
If you build AI for important decisions, you cannot hide behind novelty. You have to think about dependency, tone, overconfidence, and what happens when someone brings vulnerable context into the system. The goal is not to sound wise. It is to help people think more clearly while keeping human judgment in the center.
The user testing reinforced this more than any design principle could. The crisis detection gap was not something I would have found by reading my own code. It took someone actually typing real words that a real person in pain would type. Safety layers are not a checkbox you add at the end. They are the first thing that has to work and the last thing you should stop testing.
I started by thinking I was building a dramatic AI debate about my future. I now think I am building something more grounded: a system for people trying to move through uncertainty with more honesty, more clarity, and a better record of how they got there.
The first article was about what I built. This finalist round helped me understand what Diverge is for.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article