$60K to $1M in 7 Months: What Unlocked It for This Legal AI Founder | Devansh @ Irys

In this episode of saas.unbound, Anna Nadeina speaks with Devansh, Co-Founder and Head of AI at Irys and the writer behind the Artificial Intelligence Made Simple newsletter. The conversation moves beyond AI hype to examine what it takes to build useful infrastructure, earn enterprise trust, and make informed bets in a rapidly shifting market.

Irys grew from $60,000 in bootstrapped revenue to $1 million in seven months, while retaining every enterprise customer it signed. Devansh attributes that momentum to a combination of technical depth, product transparency, easy evaluation, and a focus on solving problems that most legal AI tools still avoid.

Why Legal AI Became the Starting Point

Devansh’s interest in legal technology began long before the current wave of generative AI. Growing up in India, he saw the consequences of an overburdened legal system: enormous case backlogs, limited resources, and people being told to return years later because their matters could not be processed in time.

That experience made legal work an obvious area for machine learning and AI. The opportunity was not simply to automate paperwork, but to help make justice and legal processes more accessible and less inefficient.

The company itself began after Devansh met his co-founder, Sabi, at a legal conference in 2024. Sabi was a BigLaw lawyer and head of innovation at a major law firm. He had access to many of the legal AI products already on the market, and found that while they could be useful in narrow situations, they often introduced new administrative work.

In other words, the effort required to use the software could cancel out much of the productivity it promised. That gap between a useful demo and a genuinely better legal workflow became the basis for Irys.

AI Signal Versus Noise: Look Beyond the RAG Wrapper

AI moves quickly enough that every week can seem like a breakthrough. New models, new agent frameworks, and new products regularly claim to transform how work gets done. But novelty is not the same as progress.

One of the central problems in the market is that many teams do not know how to build for the difficult parts of the problem. Instead, they either release products before they are confident in them or focus on easy problems and position them as technically hard ones.

Retrieval is a clear example. Many AI products are effectively “glorified RAG wrappers”: they combine vector search with a language model and present the result as a complete reasoning solution. Retrieval-augmented generation can be useful, but legal reasoning requires more than finding semantically similar chunks of text.

A more serious approach has to account for context, apply techniques conditionally, retain what has been learned, and recognize when standard retrieval will fail. The difficult parts are often left unaddressed because they are genuinely difficult.

A useful way to assess an AI product is to ask whether the team can clearly acknowledge the limitations of the current approach. A company may not have a perfect solution yet, but it should understand the hard problem it is trying to solve. If it cannot describe the limitation, it is unlikely to be solving it.

The Three Technical Bets Behind Irys

Irys made three foundational bets intended to improve legal reasoning while lowering the cost and friction of using AI systems. None are presented as guaranteed answers, but each addresses a problem that appears across industries.

1. Self-evolving memory

Legal work depends on accumulated context. Statutes can be connected, factual issues can overlap, and contradictions may need to be flagged repeatedly across a matter. Requiring an AI system to rediscover that context every time wastes tokens and risks missing important relationships.

Irys focuses on self-evolving storage and memory so that useful signals can persist. The aim is to retain learned information rather than force every new retrieval cycle to begin from scratch.

This applies beyond legal work. Coding agents, for example, can spend substantial compute relearning a codebase and its conventions. Systems that preserve usable context have the potential to become more efficient and consistent over time.

2. User steering and editable knowledge

AI systems should not treat the user as a passive recipient of outputs. Irys emphasizes user steering, allowing people to shape and evolve the knowledge structures that inform the system.

As users interact with the platform, their instructions and messages create signals that can be retained internally. This creates a more durable and personalized system, rather than one that loses important preferences and context between interactions.

The broader principle is simple: when people can articulate how information should be connected or interpreted, that input should become part of the system’s working knowledge.

3. Latent-space reasoning

The third bet is latent-space reasoning: pursuing high-quality reasoning without repeatedly encoding and decoding information through the most expensive parts of a transformer model.

The goal is greater token efficiency. If a system can reason in latent space, it may be able to achieve some of the benefits of multiple rounds of inference without repeatedly paying the full cost of those steps.

Memory, steerability, and efficient reasoning are not exclusively legal AI problems. They are common challenges in finance, programming, knowledge management, and other AI-intensive domains. That makes them stronger areas for long-term technical investment.

How to Make Better Bets in AI

No founder can know with certainty which model architecture, framework, or research direction will win. AI development is still driven by hypotheses, experiments, and informed bets.

The strongest bets tend to sit at the intersection of several industries or layers of the technology stack. A problem is more compelling when it appears in multiple contexts, such as memory challenges that affect model architectures, software agents, and hardware optimization.

This has two advantages:

• Breakthroughs in adjacent domains can create new opportunities to improve the product.
• A breakthrough in one company’s approach may have relevance across a much broader ecosystem.

These bets are riskier than incremental features, but they also carry the highest potential return. The alternative is competing for small improvements within an existing paradigm that may already be insufficient.

Building a Newsletter for Exceptional Builders, Not Everyone

Devansh’s newsletter has more than 350,000 readers, but he does not optimize it for mass appeal. His editorial philosophy is to write for a small group of exceptional builders who want access to the most important ideas at the edge of AI research.

That means much of the content is not about generic model comparisons or beginner-friendly prompts. Instead, it can cover mathematical and technical concepts such as geometry, topology, model internals, and steering within high-dimensional spaces.

The underlying mission is democratization. Compute clusters and large GPU budgets are not available to everyone, but mathematical ideas, empirical results, open-source research, and conceptual frameworks can be distributed much more broadly.

Giving more people access to these ideas can reduce the concentration of technical power. It allows capable builders outside elite institutions or conventional hiring networks to participate in the next generation of innovation.

Choose your own learning path

The volume of AI information can feel overwhelming, but overwhelm often comes from unclear priorities. The first question is not which course, model, or newsletter to follow. It is what you actually want to accomplish.

There are two valid paths:

• Focused builders can identify a problem they care deeply about and commit to studying it in detail.
• Explorers can intentionally build broad knowledge across disciplines without judging themselves against specialists.

Problems arise when someone trying to explore broadly compares their progress to someone pursuing one narrow obsession. These are different goals, and they should be measured differently.

For people who want to build in a specific area, the recommendation is to begin near the cutting edge rather than spending too long on generic foundational material. Study the history of the field, identify today’s biggest unresolved problems, and trace how those problems emerged.

Understanding why an approach worked or failed creates stronger mental models than memorizing a list of tools. Technical depth should also match individual inclination. Some people should go deep into architecture and engineering; others may be better served by understanding high-level frameworks and focusing on strategy, operations, or decision-making.

What Unlocked Irys’ Growth

Irys reached $1 million in revenue within seven months after starting with $60,000 in bootstrapped revenue. The core unlock was a deep understanding of the customer’s actual problem.

In enterprise conversations, technical credibility matters. Legal teams and their technical colleagues often ask how a product differs from a standard vector-search system, what its underlying architecture is, and why it should perform better on difficult matters.

Being able to answer those questions directly builds trust. A vague claim about custom models or proprietary orchestration is not enough when the customer is trying to understand whether the system handles reasoning differently.

Irys combines technical transparency with a product-led evaluation process. Prospective users can sign up without a credit card, access a 14-day trial, and continue on a free tier if they need more time before making a decision.

The company also encourages head-to-head testing. Rather than relying only on a polished standard demonstration, teams can test Irys against another platform using the same documents and their hardest questions. That approach lets the product prove its value in the customer’s real context.

Other growth factors include:

• Simple access to the product before a purchase commitment.
• Transparent pricing published directly on the website.
• Openness about the technical approach and research behind the system.
• Strong product outcomes that encourage organic discussion and third-party reviews.

For Irys, transparency is not merely a marketing position. It is part of the sales strategy: make it easy for potential customers to understand the product, test it, and compare it.

Workflow Products Versus Platform Players

The legal AI market can be divided broadly into workflow owners and platform players.

Workflow owners focus on a highly specific job, such as a Microsoft Word plugin or a narrow legal research workflow. These products can be easier to position and adopt because they make a clear promise: become the best tool for one particular task.

However, that focus can also create a ceiling. A company whose engineering organization, leadership, and product architecture were built around one workflow may struggle to become a general-purpose legal platform later. Meanwhile, competitors can continuously attack that workflow with their own features.

Platform players aim to become the place through which broader categories of work are completed. The upside is greater, but so is the complexity. A platform competes on many fronts at once and needs a strong underlying technical foundation to avoid spreading its efforts too thin.

The important question is not simply whether to build a workflow tool or a platform. It is whether the company has first solved the underlying problem that gives it a meaningful advantage.

For Irys, that means investing in reasoning infrastructure before prioritizing market capture for its own sake. Enterprise contracts, distribution, and integrations matter, but they do not create durable loyalty if the underlying output quality is weak.

In legal AI, network effects from customer data may also be overstated. Legal workflows vary widely, which makes it difficult to train a single general model that benefits proportionally from every additional customer. More promising personalization may come from understanding and adapting to an individual organization’s workflow rather than aggregating all data into one shared system.

Why Usage-Based Pricing Is the Natural Model for AI

Traditional SaaS pricing is built around predictable subscriptions: a fixed monthly or annual fee for access to software. AI-intensive products face a different economic reality because their costs are tied closely to model inference and token usage.

For that reason, Devansh expects AI software to move increasingly toward usage-based pricing. It is more transparent for customers and more aligned with the actual cost structure of operating token-heavy systems.

Some companies are hesitant to make that shift because inference can be expensive. If an AI task costs tens of dollars to run, purely usage-based pricing can expose both the customer and the provider to uncomfortable economics.

But avoiding the issue does not make it disappear. AI businesses that consume meaningful compute need a pricing model that can ultimately support profitable operations. Better reasoning efficiency, lower token use, and transparent measurement will become increasingly important as the market matures.

How Founders Should Implement AI in Their Companies

There is no universal “best” way to use AI at work. Different people have different working styles.

Some leaders have a very clear vision and want AI to execute it with minimal deviation. Others use AI as a collaborative thought partner to explore options, challenge assumptions, and generate possibilities. Both approaches can work.

The right approach depends on the individual and the nature of the work. For founders trying to make AI useful beyond demos, three practices stand out.

Read widely enough to understand the market

Follow AI developments through technical communities, open-source projects, discussion forums, and trending GitHub repositories. A weekly research habit can be enough to build a top-level picture of where the field is moving and where people are struggling.

Most of what you encounter will not be directly relevant. The objective is not to copy every trend. It is to understand what people are building, what they are consuming, and where gaps may exist.

Experiment to disprove your assumptions

Do not only run experiments that validate your thesis. Deliberately look for evidence that your hypothesis is wrong.

This approach produces more robust judgment. Testing specific ideas, measuring outcomes, and aggregating what succeeds and fails helps founders move from intuition to a better-grounded point of view.

As patterns emerge, it becomes easier to recognize where deeper investment is worthwhile. You do not need to become an expert in every layer of AI. You do need enough perspective to identify the few areas where your company should double down.

Build a culture that shares failures as well as wins

At an organizational level, experimentation only creates leverage if the results are shared. Teams should actively document what they tested, what went wrong, what they learned, and what they would do differently next time.

This helps prevent the same failed experiments from being repeated in different corners of the company. It also makes innovation cumulative: one person’s unsuccessful test can save another person weeks of work.

A culture that treats well-documented failure as useful knowledge will generally learn faster than one that only rewards visible successes.

What AI Changes for Technical and Non-Technical Teams

AI can reduce the wall between technical and non-technical work, but it does not erase expertise.

A person using AI to build a front end or back end more quickly does not automatically become a strong software engineer. AI can raise performance from zero toward “good enough” for many tasks, but it does not reliably create mastery.

That still represents a major change. People with deep knowledge in one domain can use AI to extend their capabilities into adjacent areas where they do not need to be specialists.

A product manager may gain a more grounded understanding of a codebase and make more informed product decisions. An engineer may use AI to better understand market context, communicate research, or reason about the broader business implications of a technical choice.

The biggest opportunity is not turning everyone into the same kind of generalist. It is helping people close important gaps, communicate more effectively across disciplines, and make better decisions with less uncertainty.

The Most Important Advice: Study the History of the Problem

The strongest takeaway for founders is to study history—not merely dates and company names, but the reasons systems became the way they are.

When a problem appears obvious, ask deeper questions:

• Why does this problem exist?
• How did the current workflow develop?
• Which previous approaches failed, and why?
• What constraints shaped the existing system?

Context turns a surface-level complaint into a real understanding of the opportunity. It reveals whether a problem is structural, whether it has already been tried, and where a new approach may have leverage.

In AI, where new tools appear daily, that historical grounding is especially valuable. It helps founders distinguish a temporary trend from a durable problem worth solving.