Co-founder and CEO at Bito. 

We build codebase intelligence for AI coding and reviews, so AI tools can reason about real systems instead of guessing. 

Before Bito, I built and scaled companies across multiple markets, including PubMatic (NASDAQ: PUBM). 

Right now, I spend most of my time thinking about AI, software systems, and why understanding codebases matters more than writing code faster.

Amar Goel

Super excited to share our first independent evaluation of Bito’s AI Architect on SWE-Bench Pro.

Working with The Context Lab, we evaluated AI Architect using the same Claude Sonnet 4.5 class agent, the same tools, and the same execution limits. The only difference was system context. 

With AI Architect enabled, the results were clear: 

• Task success rate increased by 39%
• Tasks completed 19.6% faster 
• Tool calls dropped by 25.4% 

All without any additional token or compute overhead.  

We intentionally focused this evaluation on large, real-world codebases with millions of lines of code and hundreds of files, because that’s where today’s coding agents struggle the most without system-level understanding. 

This reinforces our core belief: better outcomes don’t come from bigger models alone, but from giving AI real understanding of the system it’s working in. 

The full SWE-Bench Pro evaluation report is available below. 

Link in the comments.

Bito | Top score on SWE-Bench Pro - Deep codebase intelligence for AI coding and reviews

In 2025, AI-assisted pull requests showed around 1.7× more issues than human-written PRs, even though tests passed and reviews looked clean. This will only go up if we don’t understand the root: 

The reason is fairly straightforward once you look closely.  

AI code gen works with partial views of the codebase. 

It sees files and snippets, but it does not see how state flows across services, where invariants were added after past incidents, or which assumptions are baked into downstream consumers.  

When those relationships are missing, the generated code follows patterns that look reasonable locally and break behavior globally. 

The study pointed out that most of the issues appeared in places reviewers and tests are weakest. 

Conditional paths, retries, rollbacks, and failure handling are rarely exercised fully in unit tests, and PR reviews are scoped to diffs.  

If the system context is incomplete, both humans and tools end up approving code that only works under ideal conditions. 

This is why scaling AI coding tools inside large systems has been harder than demos suggest. Throughput goes up, but system risk quietly accumulates. 

This is the gap Bito’s AI Architect is designed to address.

If you want to learn more, I'd be happy to chat. DMs always open.

AI coding tools feel super cool on side projects, then break as soon as they're inside large, existing, enterprise codebases.  

Custom frameworks, internal conventions, fragmented repos, and years of accumulated constraints change the whole narrative. 

A LinkedIn post by Gergely Orosz from earlier last year captured it well, including a case where a top company rolled out an AI coding tool to thousands of developers and roughly half stopped using it a few months later. 

~50% ! 

The fact that it still feels true today matters, because it points to a structural gap (not a temporary tooling gap).

AI code gen breaks down when the tool cannot model the system, including dependencies, call paths, and the implicit contracts that make production software behave.  

That is where codebase intelligence becomes the difference between a demo and something you can trust in an enterprise. 

Exactly what Bito is building: The AI Architect.

Veracode’s 2025 GenAI code security report states that about 45% of AI-generated code still contains vulnerabilities.

This is one big reason AI-generated code keeps adding security debt. And that number has barely changed even as models have improved.

The underlying reason is straightforward. These models learn from public code, and public code includes a lot of insecure patterns.

When AI code gen is asked to implement common tasks, it often reproduces what it has seen most:

Raw SQL queries instead of parameterized ones.
Hardcoded secrets. Missing validation. And more..

The code compiles, looks reasonable in review, and passes basic tests, while carrying flaws that only surface later.

Security rules live across schemas, data flows, and service boundaries. When AI works without seeing those relationships, it fills gaps by pattern matching. That behavior scales insecure code faster than teams can realistically review it.

We built Bito’s AI Architect. It addresses this by building system intelligence through a knowledge graph of the full codebase, so AI code gen reflects how security actually works across the system.

🎤 This is what a customer has to say about the AI Architect:

"We’ve spent too much time fixing agent-generated code. Now most changes work in one shot because Bito’s AI Architect actually understands our services and APIs. It’s been a huge boost to Cursor."

AI Architect is available via MCP (on-prem coming soon).
We are rolling out demos on a huge scale.
DMs open if you want one.

taplio

The Battle

Amar Goel

Dan Rosenberg

Followers

Posts

Engagements

Likes

Performance ScoreAmar Goel's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Amar Goel

Dan Rosenberg

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Amar Goel's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreAmar Goel's LinkedIn Performance