IdX — Intelligence × Evidence × Execution
Inspiration
Every startup begins with a question.
“Why doesn't this exist?” “Why are people struggling with this?” “Can we build this better?” “What if we tried?”
But after that first question comes the difficult part:
Finding the right answer.
When we started thinking about this problem, we realized that founders, researchers, students, and teams already have access to enormous amounts of information. They can search the web, read research papers, analyze reports, study competitors, look through customer discussions, and use AI assistants.
Yet making a good decision is still difficult.
The problem is not simply finding information.
The problem is knowing:
- What information actually matters?
- What are we still missing?
- Which assumptions are dangerous?
- What problems are people actually facing?
- What has already failed?
- Where are the opportunities?
- What contradictions exist?
- And most importantly, what should we investigate next?
That led us to one question:
«What if AI could investigate a problem with you instead of simply answering questions about it?»
That question became IdX — Intelligence × Evidence × Execution.
What it does
IdX is an experimental evidence-driven decision intelligence system.
Instead of treating an investigation as:
Question → Answer
IdX explores:
Question → Investigation → Evidence → Unknowns → Better Question → Decision
A user can start with something as simple as:
«“I want to build an AI-powered online learning platform.”»
Instead of immediately generating a business plan, IdX is designed to first understand the case and investigate it.
It can:
- understand the user's objective and current situation,
- represent the case and its constraints,
- retrieve relevant research,
- examine discussion-based signals,
- identify potential pain points and needs,
- detect recurring factors,
- distinguish evidence sources and their verification status,
- identify gaps and unknowns,
- generate candidate questions,
- rank questions according to their potential decision value,
- update the investigation when the user answers,
- and continue the investigation as the case evolves.
The central idea is that every answer should change what the system knows about the case.
That changed understanding should influence what it asks or investigates next.
The long-term loop we envision is:
Case → Evidence → Question → Answer → Decision → Experiment → Outcome → Learning
IdX is not intended to simply tell someone what to do.
It is intended to help them understand why a decision should be made, what evidence supports it, what is still uncertain, and what should be tested next.
How we built it
We built the prototype primarily with Python and FastAPI, using a modular investigation architecture.
The system maintains a structured case state containing information such as the objective, current state, constraints, evidence, assumptions, unknowns, and investigation history.
We then built several layers around that state.
Evidence Retrieval
IdX can retrieve relevant research from a structured research corpus.
For the prototype, we intentionally use a transparent lexical retrieval baseline rather than claiming that it is already a fully semantic retrieval system.
We also experimented with discussion-style evidence so that potential user pain points can be considered alongside formal research.
Evidence Context
Different evidence sources are converted into a common context while preserving:
- source type,
- source identity,
- relevance,
- provenance,
- and verification status.
This allows IdX to distinguish between something supported by research and something that is only an unverified discussion signal.
Signal Extraction
The system extracts potential signals such as:
- complaints,
- needs,
- suggestions,
- failure factors,
- recurring factors,
- and potential opportunities.
These signals are treated as inputs for further investigation rather than automatically being declared facts.
Adaptive Question Engine
The system generates candidate questions based on the current case state.
Questions are evaluated using factors such as:
- relevance,
- information gain,
- decision impact,
- evidence value,
- uncertainty reduction,
- and user effort.
The current ranking approach is deliberately transparent and acts as a baseline for a future learned ranking model.
The architecture is also intended to remain domain-independent.
We did not want separate hard-coded question flows for agriculture, hotels, education, SaaS, or other industries.
Instead, the same investigation mechanism should adapt to the case itself.
Challenges we ran into
The hardest part was realizing that building a system that looks intelligent is much easier than building one that can be honestly called intelligent.
We experimented with machine-learning approaches for areas such as question ranking, factor discovery, and language understanding.
But we quickly encountered a fundamental problem:
Good investigation data is difficult to obtain.
To train a truly useful question-ranking model, we need trajectories showing:
Case → Evidence → Question → Answer → Decision → Outcome
We need to know not just which question was asked, but whether asking that question actually improved the eventual decision.
We also faced challenges around evidence quality.
A research paper is not automatically proof of a business opportunity.
A discussion is not automatically a market fact.
One complaint is not automatically a validated customer need.
And an AI-generated statement is not automatically evidence.
This forced us to design IdX around provenance and uncertainty instead of hiding those limitations.
Another challenge was resisting the temptation to hard-code intelligence into individual domains. We wanted the system to investigate different problems using the same underlying reasoning framework.
Accomplishments that we're proud of
We are proud that we moved IdX beyond the idea stage and built a working investigation pipeline.
The current prototype can take a case, retrieve multiple forms of evidence, combine them into a bounded evidence context, extract signals, maintain investigation state, and generate an adaptive next question.
For example, in our education demonstration, IdX combined research evidence with discussion signals around personalized and adaptive learning.
The important part was not simply retrieving those documents.
It was creating a pipeline where the evidence becomes part of the investigation and can influence what the system asks next.
We are also proud of the architectural decision to keep the system domain-independent.
Instead of building dozens of predefined decision trees, we built a general case representation and investigation loop.
Most importantly, we are proud of being technically honest about what we have built.
The current system is a research prototype.
It does not pretend to have solved general decision intelligence.
The transparent baselines we built give us a foundation that can eventually be replaced with stronger semantic retrieval, learned question ranking, evidence synthesis, contradiction detection, and outcome-driven learning.
What we learned
The biggest lesson was simple:
Intelligence is not the same as information.
A system can retrieve hundreds of documents and still fail to help someone make a decision.
The difficult part is understanding which information changes the decision.
We also learned that asking questions is much harder than generating questions.
Generating ten reasonable questions is easy.
Choosing the one question that is most valuable given everything currently known is a much harder problem.
We learned that evidence needs context.
Different sources have different reliability, and the system must preserve that distinction.
Most importantly, we learned that real decision intelligence requires feedback.
A question cannot truly be judged only by whether it sounds relevant.
Eventually, we need to know:
Did the answer change the decision?
Did the decision lead to a better outcome?
That is the learning signal we ultimately want IdX to capture.
What's next for IdX
The current prototype is only the beginning.
Our next goal is to make IdX substantially better at understanding and investigating real-world cases.
We want to move toward:
- hybrid and semantic evidence retrieval,
- larger and more diverse public evidence sources,
- stronger evidence synthesis,
- contradiction detection,
- better factor discovery,
- uncertainty and assumption tracking,
- learned question ranking,
- decision-impact estimation,
- experiment design,
- and outcome-based learning.
The most important long-term dataset we want to build is not simply a collection of documents.
It is a collection of investigation trajectories:
Case → Evidence → Question → Answer → Decision → Experiment → Outcome
With enough high-quality trajectories, IdX could learn which questions actually reduce uncertainty and improve decisions.
Ultimately, we want IdX to become a system that doesn't just answer:
«“What should I do?”»
but helps a person discover:
«“What do I need to understand before I can know what I should do?”»
Because sometimes the most valuable intelligence isn't the answer you receive.
It is discovering that you were asking the wrong question in the first place.
That is why we built IdX.
Log in or sign up for Devpost to join the conversation.