1. The Question in the Dark

I have had countless days with discarded notebook drawings and glowing monitors in the dark of the night, staring at a blank page or a screen, asking a question that felt both ancient and urgent: How do we actually learn?

Not how we cram for a test, skim a summary, or copy-paste code from a search engine—but how a human mind truly transforms. How do we learn when we are completely alone in a quiet room? How do we learn when we sit across from another human being? And what happens when we introduce Artificial Intelligence into that equation—not to replace our intelligence, but to help us understand how our own minds work?

When I first set out to build ProfilEd, I wasn’t trying to build a startup product or a flashcard app. I was driven by a deep, personal obsession to build a tool for human amplification. I wanted to build something that would teach people how to think, how to reason, and how to hold their ground through complex scenarios across every aspect of life.

That pursuit forced me out of modern software paradigms and sent me tumbling backward into cognitive history. I found myself spending nights poring over the works of Lev Vygotsky, John Sweller, and Benjamin Bloom. I was reading Vygotsky’s thoughts on the Zone of Proximal Development—that delicate, magical threshold where a learner is pushed just enough to grow without snapping. I was studying Sweller’s Cognitive Load Theory, trying to understand why our mental bandwidth freezes up when information is presented the wrong way.

I realized that if AI is used simply as a quick-answer engine, it acts as an intellectual crutch. It makes us passive. But if it could be built to hold up a mirror to our reasoning, it could trigger an awakening.

In an era where every headline promised an AI tool to write your essays or generate your answers, I set out to build something counter-cultural. I didn't want to build an information vending machine. I wanted to build a catalyst for human amplification—a platform that would teach people how to think, how to reason, and how to navigate the messy, high-stakes complexity of modern life.

That quest became ProfilEd.

What began as a software project quickly turned into a profound personal journey. To build ProfilEd, I had to walk away from conventional tech frameworks and dive headlong into cognitive theory. I spent nights poring over Lev Vygotsky’s Zone of Proximal Development (ZPD), John Sweller’s Cognitive Load Theory, and Benjamin Bloom’s Taxonomy.

I realized that if an AI system simply gives away the answer, it robs the human brain of the essential "desirable difficulty" required for neural consolidation. But if an AI system can be engineered as a cognitive mirror—reflecting your reasoning gaps, testing your boundaries, and dynamically simulating reality—it can unlock an extraordinary evolution in how we learn alone, how we learn together, and how we learn alongside artificial intelligence.

This is the story of how ProfilEd was built, the unexpected breakthroughs and technical battles along the way, and what a year of building with AI taught me about the human mind.

2. The Aspirant’s Gauntlet and the Relentless Mirror

To understand what a learner goes through, I decided to become the subject of my own experiments. I spent months putting myself through intense, simulated study sessions, pretending to be aspirants in fields I wanted to deeply understand.

One night, I would assume the persona of an ambitious candidate preparing for a McKinsey case interview, trying to structure messy, ambiguous business problems into clean, mutually exclusive frameworks. The next night, I would dive into the strange, counter-intuitive world of quantum computing, wrestling with Hilbert spaces, superposition, and qubit state vectors. On another, I was stepping into the shoes of a FAANG Product Manager designing large-scale distributed systems, or trying to unpack what Nvidia's Jensen Huang really meant when he declared that "AGI is here."

Episode 1: The missing branch in my framework

One night at 2:00 AM, I pretended to be a candidate preparing for a partner-level McKinsey interview. I fed the system a complex Business Analysis meeting scenario involving a regional energy provider attempting a digital transition.

I attempted to outline a framework using standard business jargon, jumping smoothly from market sizing to financial projections. But ProfilEd didn't smile and nod. It paused. It isolated a missing branch in my logic tree, highlighting that my market analysis failed to separate regulatory subsidies from organic consumer demand—violating the MECE (Mutually Exclusive, Collectively Exhaustive) principle.

I had not read about McKinsey MECE principle in a long time. It was "Go back to the Basics" moment. It meant something as simple as "So What?" can totally change the game. Ever since the dawn of AI was I had stopped asking these questions. In late 2021, when everyone got exposed to AI, the way we access information completely changed and drifted to our own created echo chambers to start with; and later in 2026, we are learning and improving things at scale of information that someone as average as me is struggling to work with. It is not that I am not used to volumes of data and insight, I have been an information hoarder and a complete bookworm looking for gold almost all the time. I had once downloaded TBs of books so that I could just glance through them once. I have read countless research papers; half of which I wouldn't even completely understand. So I think I can handle dense information and sift through it bit by bit by asking questions; looking for answers on Google.

Yet here I was, no matter what I knew, with these special scaffolds of prompts the AI caught me off guard, questioned me well; and left me thinking.

The AI didn't just tell me I was wrong; it forced me into a Socratic corner: "You’ve assumed organic adoption, but regulatory tariffs in Region B shift next quarter. How does your cash-flow model hold up if subsidies drop by 40%?"

I was stunned. I sat back in my chair, sweating through the math. For the first time, I felt the uncanny power of an AI system that refused to be dazzled by rhetoric.

I could only say to myself - "What?"

Episode 2: Decoding Jensen Huang’s "AGI is Here"

When Nvidia's Jensen Huang famously declared that "AGI is here", depending on how you define intelligence, I sat down with ProfilEd to unpack the statement. We didn't just summarize news articles. We broke down transformer compute scaling laws, FLOP efficiency curves, and memory bandwidth bottlenecks versus human synaptic density.

The whole conversation I just sat back and enjoyed passively between 3 or 4 AIs; it was fascinating how it explained how much wisdom was packed into this simple statement; and also how much it could be just human oversight. Nobody knows the future. What are the rules of the game? What actually is AGI.

I remember having several study sessions to see AIs debate on it from several perspectives. Infact in one of the study sessions the digital twin of Jensen Huang himself was there to debate with Demis Hassabis.

Boy it was fun!! And not just that I promise you. It was so much more!

When I was testing our realtime AI interviews; the AI interviewers I noticed would switch and start talking in different languages if the conversation grew too long. Those were early days of tokens. I noticed how Deepseek would just start repeating stuff.

And similarly there were some really thought-provoking, grounding conversations and AI interviews to see how little I knew.

During these sessions, I subjected myself to hundreds of live AI-driven interviews. I would sit back, just role-play and see how much I could learn from them; about me. It was BANG ON!!

And then came the humbling moment that changed everything: I realized I couldn't manipulate the AI beyond a point.

When I couldn't fool it, I created a digital twin of mine and created a shadow AI interview where my digital twin would do that same exact interview with the AI. I saw how it managed to use the famous STAR methodology—Situation, Task, Action, Result—to frame its answers.

In real life, when you talk to another human, you can often rely on charisma, fluid language, or confident posture to gloss over a weak argument. But the AI didn't care about my tone. It caught every single hand-wave. It picked up on the subtle gap where I skipped from Situation straight to Result without explaining the structural Action in between. It called out my hand-waving when my framework wasn't mutually exclusive.

It was an uncomfortable, beautiful realization: the machine had become an unsparing mirror. It wasn't just testing what I knew; it was exposing the exact boundary where my confidence exceeded my actual understanding.

3. Ghosts in the Machine: Cache Poisoning and the Unified AI Service

Building with AI teaches you a lot no matter who you are; whether a seasoned expert or a complete newbie. During these conversational exchanges I also got to witness several ghosts in the machine; each with its own challenges, biases and strengths.

As I built out the engine, I was working with a choir of different AI models—Claude, Gemini, Qwen, DeepSeek, Nemotron, and others. I thought that by combining these different cognitive architectures, I would get the ultimate synthetic mind.

Instead, I unlocked a chaotic, fascinating mess.

I began to see models drift. When you pass dynamic conversation histories back and forth across different model APIs, something strange happens to their memory: cache poisoning. One model would inject a subtle structural assumption or tone into the transcript, and by the third turn, another model would absorb that assumption, lose its own identity, and start hallucinating false confidence. That gave birth to our AI council Consensus Framework. Simply put it was adversarial debate to put things in perspective, to validate the truths, to prioritize the evident facts, to think about information asymmetry.

Worse still, I saw how stale data and internet bias embedded deep within model weights would leak into conversations. The models would subtly steer the user toward conventional, outdated wisdom instead of encouraging first-principles thinking.

I didn't design the Unified AI Service because I wanted a clean architectural diagram. I built it out of pure survival. I needed a central intelligence layer that could route queries based on information intent, de-poison context windows in real time, ground the models against stale internet bias, and keep each AI strictly within its intended pedagogical bounds.

4. The Privacy-Preserving Digital Twin

As I watched people interact with early prototypes, another problem became clear: self-reported learning data is almost always wrong. If you ask someone what they want to learn, they describe their ideal self. If you ask them how well they understand a concept, imposter syndrome or overconfidence skews the answer.

Some of the early users of ProfilEd helped understand their resistance with the technology and whole UX of the application while it was building. Some did not talk. They were shy and were mostly passive observers. Some said our study sessions where multiple AIs debate and talk is a podcast to watch.

Some gave me a feedback that conversation is too restricted or too fluid or too dense. I had to solve very interesting problems which helped me understand the nuance of such information and how it explodes.

While implementing those I had to jump very deep into understanding Godel limit or Kolmogorov complexity or Shannon Entropy. These were completely new and advanced concepts for me. The entropy of any conversation, what kind of basin it is and how does it change and drift. It was mind boggling. I have captured some of my learnings in our whitepapers; that I intend to soon publish.

I wanted to capture the true, living footprint of how a person learns—their pauses, their hesitations, their sudden bursts of insight—without ever invading their space or violating their privacy.

We architected a privacy-first model that complies strictly with global Data Privacy Laws. Instead of collecting intrusive personal data, the system creates an anonymized Digital Twin—a dynamic model representing the person the user is trying to become.

By observing non-intrusive temporal signals—how long someone hesitates before typing a response, where their attention blurs, how quickly they self-correct during a Socratic dialogue—ProfilEd constructs a 300-Dimension Behavioral DNA.

Infact I built a complete mind and brain architecture driven mapping. I have copied some screenshots of these representations for you below.

It allowed us to pinpoint the "question in hiding": the fundamental prerequisite gap that a user doesn't even realize they have. It showed us when a user was suffering from cognitive overload (Sweller) versus when they were simply bored, allowing the system to adjust the learning difficulty on the fly.

5. The Pedagogy of Resistance: Teaching How to Fish

The easiest thing an AI can do is give you the answer. It is also the most destructive thing it can do for your learning.

When an AI hands you a ready-made solution, your brain registers relief, but zero neural connections are forged. I realized that to help people learn, the AI tutor had to practice the pedagogy of resistance. It had to learn when not to answer. It had to learn how to teach a person how to fish.

This realization led to the creation of the MACI Score (Multi-Agent Cognitive Index).

The MACI score came out of a moment of frustration when watching someone recite a memorized answer with supreme overconfidence. I wondered: What if we subjected their explanation to a panel of multiple AI models, each operating at peak domain proficiency, pushing back from different angles simultaneously?

The MACI score measures how many distinct AI models—and at what level of cognitive rigor—a user can satisfy and convince through original Socratic defense. You cannot fake a MACI score. You cannot copy-paste your way through it. You have to understand the material so deeply that your logic holds up under cross-examination.

At the same time, we realized that forcing people to type out long essays created massive cognitive fatigue. Learning shouldn't feel like filling out tax forms. We shifted toward low-latency, voice-first interactions and three modes of engagement:

1. Observer Mode: Sitting back and watching multiple AI agents debate complex topics, allowing users to absorb high-level reasoning passively before jumping in.
2. Real-Time Voice Sparring: Talking through complex scenarios naturally, feeling the immediate, conversational give-and-take of a live panel.
3. Adaptive Simulations: Turning static ideas into living, unpredictable scenarios where the user has to make decisions under pressure.

Here is a quick sneak-peek into 3 very core human + AI interactions in Profiled and how they work.

6. Gratitude and What a Year Taught Me

Looking back over this year of building, testing, failing, and iterating, I am filled with a deep sense of gratitude for the people who unknowingly helped shape this vision.

These expeditions with AI have of course come at a sincere cost of time, money and relationships. Now, looking back, I see it was a heavy price to pay; or may be not. I am still contemplating. Everyone around me has been using AI to learn in their own way, creating their own echo chambers, sifting through their own thoughts, victimized and subject to the noise of other first movers.

We are living in the most excitingly scary times. Past two years of this certain kind of a struggle to learn more, has also taught me a lot even while I was away from screen - just through simple human conversations.

I think of the endless late-night study sessions with friends, colleagues, and aspirants who volunteered to be test subjects—letting an early, unpolished system probe their thinking and expose their knowledge gaps.

And I also think of the quiet conversations, like the one I had with senior banking executives who were trying to figure out how a massive, highly-regulated institution could adopt AI without losing human judgment or compromising safety. I met them at AI summit. One thing is clear - everyone is trying to figure something out.

Whether it is the Professors in Quantum Computing, who are active researchers and are debunking myths and teaching the future generation of experts and scientists. Everyone has a question. Everyone has a dark territory of known unknowns and unknown unknowns mapped out.

And fortunately, I got a chance to have some of these interesting insightful interactions in different geographies both in the East and West Coast of the US. Both in the eastern and the western part of the world. I have felt super confident, super confused, super excited and keen and also very fearful at times. Yet, it was rewarding in many ways.

I am very thankful for some of these conversations.

Watching them wrestle with real-world risk, compliance, and governance directly inspired how we built scenario-based learning and Observer Mode.

If building ProfilEd taught me one ultimate truth, it is this: The future of AI in education is not about building smarter machines to think for us. It is about building tools that inspire us to think deeper for ourselves.

AI should not be a crutch that makes human minds passive; it should be an intellectual catalyst that expands our Zone of Proximal Development, tears down our overconfidence, respects our privacy, and reveals the extraordinary capacity for reasoning that lives inside every one of us.

We built ProfilEd to give people a tool to learn how to fish, how to reason, and how to navigate an increasingly complex world. And after a year of walking this path, I am more convinced than ever that when you give a human being the right mirror, there is no limit to what they can learn.

Written from personal notes, architectural journals, and late-night study sessions over a year of building ProfilEd.

The Architecture of Awakening: A Year of Learning How Humans Learn Alongside AI

1. The Question in the Dark

I have had countless days with discarded notebook drawings and glowing monitors in the dark of the night, staring at a blank page or a screen, asking a question that felt both ancient and urgent: How do we actually learn?

Not how we cram for a test, skim a summary, or copy-paste code from a search engine—but how a human mind truly transforms. How do we learn when we are completely alone in a quiet room? How do we learn when we sit across from another human being? And what happens when we introduce Artificial Intelligence into that equation—not to replace our intelligence, but to help us understand how our own minds work?

When I first set out to build ProfilEd, I wasn’t trying to build a startup product or a flashcard app. I was driven by a deep, personal obsession to build a tool for human amplification. I wanted to build something that would teach people how to think, how to reason, and how to hold their ground through complex scenarios across every aspect of life.

That pursuit forced me out of modern software paradigms and sent me tumbling backward into cognitive history. I found myself spending nights poring over the works of Lev Vygotsky, John Sweller, and Benjamin Bloom. I was reading Vygotsky’s thoughts on the Zone of Proximal Development—that delicate, magical threshold where a learner is pushed just enough to grow without snapping. I was studying Sweller’s Cognitive Load Theory, trying to understand why our mental bandwidth freezes up when information is presented the wrong way.

I realized that if AI is used simply as a quick-answer engine, it acts as an intellectual crutch. It makes us passive. But if it could be built to hold up a mirror to our reasoning, it could trigger an awakening.

In an era where every headline promised an AI tool to write your essays or generate your answers, I set out to build something counter-cultural. I didn't want to build an information vending machine. I wanted to build a catalyst for human amplification—a platform that would teach people how to think, how to reason, and how to navigate the messy, high-stakes complexity of modern life.

That quest became ProfilEd.

What began as a software project quickly turned into a profound personal journey. To build ProfilEd, I had to walk away from conventional tech frameworks and dive headlong into cognitive theory. I spent nights poring over Lev Vygotsky’s Zone of Proximal Development (ZPD), John Sweller’s Cognitive Load Theory, and Benjamin Bloom’s Taxonomy.

I realized that if an AI system simply gives away the answer, it robs the human brain of the essential "desirable difficulty" required for neural consolidation. But if an AI system can be engineered as a cognitive mirror—reflecting your reasoning gaps, testing your boundaries, and dynamically simulating reality—it can unlock an extraordinary evolution in how we learn alone, how we learn together, and how we learn alongside artificial intelligence.

This is the story of how ProfilEd was built, the unexpected breakthroughs and technical battles along the way, and what a year of building with AI taught me about the human mind.

2. The Aspirant’s Gauntlet and the Relentless Mirror

To understand what a learner goes through, I decided to become the subject of my own experiments. I spent months putting myself through intense, simulated study sessions, pretending to be aspirants in fields I wanted to deeply understand.

One night, I would assume the persona of an ambitious candidate preparing for a McKinsey case interview, trying to structure messy, ambiguous business problems into clean, mutually exclusive frameworks. The next night, I would dive into the strange, counter-intuitive world of quantum computing, wrestling with Hilbert spaces, superposition, and qubit state vectors. On another, I was stepping into the shoes of a FAANG Product Manager designing large-scale distributed systems, or trying to unpack what Nvidia's Jensen Huang really meant when he declared that "AGI is here."

Episode 1: The missing branch in my framework

One night at 2:00 AM, I pretended to be a candidate preparing for a partner-level McKinsey interview. I fed the system a complex Business Analysis meeting scenario involving a regional energy provider attempting a digital transition.

I attempted to outline a framework using standard business jargon, jumping smoothly from market sizing to financial projections. But ProfilEd didn't smile and nod. It paused. It isolated a missing branch in my logic tree, highlighting that my market analysis failed to separate regulatory subsidies from organic consumer demand—violating the MECE (Mutually Exclusive, Collectively Exhaustive) principle.

I had not read about McKinsey MECE principle in a long time. It was "Go back to the Basics" moment. It meant something as simple as "So What?" can totally change the game. Ever since the dawn of AI was I had stopped asking these questions. In late 2021, when everyone got exposed to AI, the way we access information completely changed and drifted to our own created echo chambers to start with; and later in 2026, we are learning and improving things at scale of information that someone as average as me is struggling to work with. It is not that I am not used to volumes of data and insight, I have been an information hoarder and a complete bookworm looking for gold almost all the time. I had once downloaded TBs of books so that I could just glance through them once. I have read countless research papers; half of which I wouldn't even completely understand. So I think I can handle dense information and sift through it bit by bit by asking questions; looking for answers on Google.

Yet here I was, no matter what I knew, with these special scaffolds of prompts the AI caught me off guard, questioned me well; and left me thinking.

The AI didn't just tell me I was wrong; it forced me into a Socratic corner: "You’ve assumed organic adoption, but regulatory tariffs in Region B shift next quarter. How does your cash-flow model hold up if subsidies drop by 40%?"

I was stunned. I sat back in my chair, sweating through the math. For the first time, I felt the uncanny power of an AI system that refused to be dazzled by rhetoric.

I could only say to myself - "What?"

Episode 2: Decoding Jensen Huang’s "AGI is Here"

When Nvidia's Jensen Huang famously declared that "AGI is here", depending on how you define intelligence, I sat down with ProfilEd to unpack the statement. We didn't just summarize news articles. We broke down transformer compute scaling laws, FLOP efficiency curves, and memory bandwidth bottlenecks versus human synaptic density.

The whole conversation I just sat back and enjoyed passively between 3 or 4 AIs; it was fascinating how it explained how much wisdom was packed into this simple statement; and also how much it could be just human oversight. Nobody knows the future. What are the rules of the game? What actually is AGI.

I remember having several study sessions to see AIs debate on it from several perspectives. Infact in one of the study sessions the digital twin of Jensen Huang himself was there to debate with Demis Hassabis.

Boy it was fun!! And not just that I promise you. It was so much more!

When I was testing our realtime AI interviews; the AI interviewers I noticed would switch and start talking in different languages if the conversation grew too long. Those were early days of tokens. I noticed how Deepseek would just start repeating stuff.

And similarly there were some really thought-provoking, grounding conversations and AI interviews to see how little I knew.

During these sessions, I subjected myself to hundreds of live AI-driven interviews. I would sit back, just role-play and see how much I could learn from them; about me. It was BANG ON!!

And then came the humbling moment that changed everything: I realized I couldn't manipulate the AI beyond a point.

When I couldn't fool it, I created a digital twin of mine and created a shadow AI interview where my digital twin would do that same exact interview with the AI. I saw how it managed to use the famous STAR methodology—Situation, Task, Action, Result—to frame its answers.

In real life, when you talk to another human, you can often rely on charisma, fluid language, or confident posture to gloss over a weak argument. But the AI didn't care about my tone. It caught every single hand-wave. It picked up on the subtle gap where I skipped from Situation straight to Result without explaining the structural Action in between. It called out my hand-waving when my framework wasn't mutually exclusive.

It was an uncomfortable, beautiful realization: the machine had become an unsparing mirror. It wasn't just testing what I knew; it was exposing the exact boundary where my confidence exceeded my actual understanding.

3. Ghosts in the Machine: Cache Poisoning and the Unified AI Service

Building with AI teaches you a lot no matter who you are; whether a seasoned expert or a complete newbie. During these conversational exchanges I also got to witness several ghosts in the machine; each with its own challenges, biases and strengths.

As I built out the engine, I was working with a choir of different AI models—Claude, Gemini, Qwen, DeepSeek, Nemotron, and others. I thought that by combining these different cognitive architectures, I would get the ultimate synthetic mind.

Instead, I unlocked a chaotic, fascinating mess.

I began to see models drift. When you pass dynamic conversation histories back and forth across different model APIs, something strange happens to their memory: cache poisoning. One model would inject a subtle structural assumption or tone into the transcript, and by the third turn, another model would absorb that assumption, lose its own identity, and start hallucinating false confidence. That gave birth to our AI council Consensus Framework. Simply put it was adversarial debate to put things in perspective, to validate the truths, to prioritize the evident facts, to think about information asymmetry.

Worse still, I saw how stale data and internet bias embedded deep within model weights would leak into conversations. The models would subtly steer the user toward conventional, outdated wisdom instead of encouraging first-principles thinking.

I didn't design the Unified AI Service because I wanted a clean architectural diagram. I built it out of pure survival. I needed a central intelligence layer that could route queries based on information intent, de-poison context windows in real time, ground the models against stale internet bias, and keep each AI strictly within its intended pedagogical bounds.

4. The Privacy-Preserving Digital Twin

As I watched people interact with early prototypes, another problem became clear: self-reported learning data is almost always wrong. If you ask someone what they want to learn, they describe their ideal self. If you ask them how well they understand a concept, imposter syndrome or overconfidence skews the answer.

Some of the early users of ProfilEd helped understand their resistance with the technology and whole UX of the application while it was building. Some did not talk. They were shy and were mostly passive observers. Some said our study sessions where multiple AIs debate and talk is a podcast to watch.

Some gave me a feedback that conversation is too restricted or too fluid or too dense. I had to solve very interesting problems which helped me understand the nuance of such information and how it explodes.

While implementing those I had to jump very deep into understanding Godel limit or Kolmogorov complexity or Shannon Entropy. These were completely new and advanced concepts for me. The entropy of any conversation, what kind of basin it is and how does it change and drift. It was mind boggling. I have captured some of my learnings in our whitepapers; that I intend to soon publish.

I wanted to capture the true, living footprint of how a person learns—their pauses, their hesitations, their sudden bursts of insight—without ever invading their space or violating their privacy.

We architected a privacy-first model that complies strictly with global Data Privacy Laws. Instead of collecting intrusive personal data, the system creates an anonymized Digital Twin—a dynamic model representing the person the user is trying to become.

By observing non-intrusive temporal signals—how long someone hesitates before typing a response, where their attention blurs, how quickly they self-correct during a Socratic dialogue—ProfilEd constructs a 300-Dimension Behavioral DNA.

Infact I built a complete mind and brain architecture driven mapping. I have copied some screenshots of these representations for you below.

It allowed us to pinpoint the "question in hiding": the fundamental prerequisite gap that a user doesn't even realize they have. It showed us when a user was suffering from cognitive overload (Sweller) versus when they were simply bored, allowing the system to adjust the learning difficulty on the fly.

5. The Pedagogy of Resistance: Teaching How to Fish

The easiest thing an AI can do is give you the answer. It is also the most destructive thing it can do for your learning.

When an AI hands you a ready-made solution, your brain registers relief, but zero neural connections are forged. I realized that to help people learn, the AI tutor had to practice the pedagogy of resistance. It had to learn when not to answer. It had to learn how to teach a person how to fish.

This realization led to the creation of the MACI Score (Multi-Agent Cognitive Index).

The MACI score came out of a moment of frustration when watching someone recite a memorized answer with supreme overconfidence. I wondered: What if we subjected their explanation to a panel of multiple AI models, each operating at peak domain proficiency, pushing back from different angles simultaneously?

The MACI score measures how many distinct AI models—and at what level of cognitive rigor—a user can satisfy and convince through original Socratic defense. You cannot fake a MACI score. You cannot copy-paste your way through it. You have to understand the material so deeply that your logic holds up under cross-examination.

At the same time, we realized that forcing people to type out long essays created massive cognitive fatigue. Learning shouldn't feel like filling out tax forms. We shifted toward low-latency, voice-first interactions and three modes of engagement:

1. Observer Mode: Sitting back and watching multiple AI agents debate complex topics, allowing users to absorb high-level reasoning passively before jumping in.
2. Real-Time Voice Sparring: Talking through complex scenarios naturally, feeling the immediate, conversational give-and-take of a live panel.
3. Adaptive Simulations: Turning static ideas into living, unpredictable scenarios where the user has to make decisions under pressure.

Here is a quick sneak-peek into 3 very core human + AI interactions in Profiled and how they work.

6. Gratitude and What a Year Taught Me

Looking back over this year of building, testing, failing, and iterating, I am filled with a deep sense of gratitude for the people who unknowingly helped shape this vision.

These expeditions with AI have of course come at a sincere cost of time, money and relationships. Now, looking back, I see it was a heavy price to pay; or may be not. I am still contemplating. Everyone around me has been using AI to learn in their own way, creating their own echo chambers, sifting through their own thoughts, victimized and subject to the noise of other first movers.

We are living in the most excitingly scary times. Past two years of this certain kind of a struggle to learn more, has also taught me a lot even while I was away from screen - just through simple human conversations.

I think of the endless late-night study sessions with friends, colleagues, and aspirants who volunteered to be test subjects—letting an early, unpolished system probe their thinking and expose their knowledge gaps.

And I also think of the quiet conversations, like the one I had with senior banking executives who were trying to figure out how a massive, highly-regulated institution could adopt AI without losing human judgment or compromising safety. I met them at AI summit. One thing is clear - everyone is trying to figure something out.

Whether it is the Professors in Quantum Computing, who are active researchers and are debunking myths and teaching the future generation of experts and scientists. Everyone has a question. Everyone has a dark territory of known unknowns and unknown unknowns mapped out.

And fortunately, I got a chance to have some of these interesting insightful interactions in different geographies both in the East and West Coast of the US. Both in the eastern and the western part of the world. I have felt super confident, super confused, super excited and keen and also very fearful at times. Yet, it was rewarding in many ways.

I am very thankful for some of these conversations.

Watching them wrestle with real-world risk, compliance, and governance directly inspired how we built scenario-based learning and Observer Mode.

If building ProfilEd taught me one ultimate truth, it is this: The future of AI in education is not about building smarter machines to think for us. It is about building tools that inspire us to think deeper for ourselves.

AI should not be a crutch that makes human minds passive; it should be an intellectual catalyst that expands our Zone of Proximal Development, tears down our overconfidence, respects our privacy, and reveals the extraordinary capacity for reasoning that lives inside every one of us.

We built ProfilEd to give people a tool to learn how to fish, how to reason, and how to navigate an increasingly complex world. And after a year of walking this path, I am more convinced than ever that when you give a human being the right mirror, there is no limit to what they can learn.

Written from personal notes, architectural journals, and late-night study sessions over a year of building ProfilEd.