Noah Labs

Noah Labs triples accuracy per dollar by modernizing legacy software on Google Cloud

Results on Google Cloud
  • 3x higher accuracy per dollar for code modernization

  • 3.1x to 7.5x lower evaluation spend across teams

  • 97.6% coding accuracy (#1 across 6 tested models)

  • 65.6% faster response speed for developer workflows

Noah Labs uses Gemma 4 on Google Cloud to help software teams modernize legacy code with 3x higher accuracy per dollar.

The high cost of model bloat

Palo Alto-based Noah Labs builds AI tools for defense contractors, government agencies, banks, and enterprise teams. The company provides mission-critical software and autonomous systems designed to modernize legacy codebases and monitor navigation reliability. When organizations set out to refactor complex, aging codebases into clean, maintainable systems, their engineers need AI systems that fit within fast-moving development cycles. A practical engineering system must assemble context across codebases, use development tools, run automated verification checks on proposed changes, and present clear evidence for human review.

Standard industry practice assumes that deploying larger frontier models would automatically produce superior engineering outcomes. However, relying on oversized models can introduce sluggish response times that break developer concentration and drive up infrastructure bills. For developers moving through continuous iteration cycles—modifying code, running tests, inspecting results, and refining—waiting on slow token generation disrupts focus and stalls deployment pipelines. These infrastructure demands made continuous, team-wide AI adoption cost-prohibitive.

Noah Labs recognized that real-world productivity is not won on synthetic leaderboard rankings. While building Sentinel, its long-context software-engineering system, the team focused its criteria on business value—seeing how much completed, validated engineering work a system can deliver for every dollar and every second of operation.

For Sentinel, we wanted the strongest operating point across quality, throughput, and cost. For developers, speed matters across the entire loop: make a change, run the checks, inspect the result, and try again. Faster generation and lower GPU costs make those iterations more responsive and bring everyday, team-wide access within reach.

Murat Can Isik

Co-founder and CEO, Noah Labs

"For Sentinel, we wanted the strongest operating point across quality, throughput, and cost,” explains Murat Can Isik, Co-founder and CEO at Noah Labs. “For developers, speed matters across the entire loop: make a change, run the checks, inspect the result, and try again. Faster generation and lower GPU costs make those iterations more responsive and bring everyday, team-wide access within reach."

To support deployment across cloud and customer-controlled environments, Noah Labs needed a balanced AI foundation pairing precise reasoning with fast throughput and predictable operating costs.

Powering interactive refactoring at scale

To turn raw intelligence into dependable code refactoring, Noah Labs built Sentinel as a structured orchestration framework around the core AI model. Rather than leaving an AI model to navigate complex legacy repositories unguided, the Sentinel harness manages five critical stages of the engineering workflow: handling context retrieval across massive code bases, invoking development tools, orchestrating multi-step tasks, running automated verification on code changes, and conducting structured reviews.

Scale is one lever. The harness determines how much of that capability becomes usable work. The shift in our thinking came when we focused on completed, validated work rather than model size. Seeing a dense 31B model compete so effectively inside Sentinel reinforced how much the surrounding harness, context, tools, and feedback matter.

Berk Yilmaz

Co-founder and CTO, Noah Labs

By surrounding the model with proper context, tools, and automated feedback loops, Sentinel helps verify that code modifications are thoroughly checked before an engineer ever spends time reviewing them. "Scale is one lever. The harness determines how much of that capability becomes usable work,” says Berk Yilmaz, Co-founder and CTO at Noah Labs. “The shift in our thinking came when we focused on completed, validated work rather than model size. Seeing a dense 31B model compete so effectively inside Sentinel reinforced how much the surrounding harness, context, tools, and feedback matter."

Recognizing that single-prompt tests cannot capture how a system performs during real-world code refactoring and verification, Noah Labs evaluated its complete workflow across six different open AI models while keeping the Sentinel harness strictly constant.

Running its production and evaluation workloads on Google Cloud Compute Engine alongside Gemini platform models allowed the team to measure what matters most to engineering leaders: successful task completion, responsiveness, operating cost, and the effort required for human engineers to review the output.

Through this system-level evaluation, Google DeepMind's dense Gemma 4 31B model emerged as the clear choice for powering Sentinel. Operating inside the Sentinel harness, the compact model outperformed alternatives nearly four times its size, delivering 65.6% faster response speeds than a 120-billion-parameter alternative and nearly doubling the speed of other tested options. The sustained throughput keeps feedback loops tight, enabling software developers to iterate, test, and ship code without disruptive waiting periods.

Engineering rigor for mission-critical systems

Powering Sentinel with Gemma 4 31B on Google Cloud has delivered immediate economic and operational benefits for Noah Labs and the software engineering teams it supports. Across the full benchmark, Noah Labs reduced its evaluation compute spend by 3.1x to 7.5x compared to larger alternative models.

When measuring aggregate accuracy against total spend, Sentinel delivered 6.082 accuracy points per dollar—yielding 3.0x the accuracy per dollar of GPT-OSS-120B and 8.8x that of Granite-4.1. Noah Labs found that lower compute costs did not compromise engineering precision.

For teams maintaining mission-critical software and autonomous systems, this significant reduction in operating cost preserved high engineering quality across evaluated benchmarks. Inside the Sentinel harness, Gemma 4 31B finished first across six evaluated models on HumanEval (97.6%), MATH-500 (93.8%), and GSM8K (97.0%), while reaching a 168.4 useful-work-per-second score.

Benchmark the complete workflow on your own tasks before choosing a model. Measure successful completion, latency, cost, and the effort needed to review the output. With Google DeepMind’s Gemma 4 31B and Google Cloud, Noah Labs is advancing practical AI for legacy software modernization, helping engineering teams turn complex codebases into understandable, testable, and maintainable systems.

Murat Can Isik

Co-founder and CEO, Noah Labs

“Noah Labs is advancing practical AI for legacy software modernization, helping engineering teams turn complex codebases into understandable, testable, and maintainable systems," says Isik. Software teams modernizing complex legacy codebases now have more room to iterate, test, and validate code changes within their existing budgets.

Leveraging Google Cloud enabled Noah Labs to eliminate the traditional trade-off between reasoning depth and infrastructure overhead, proving that lean hardware configurations can outperform heavy models. Establishing this practical operating point has fundamentally expanded how Noah Labs delivers Sentinel to enterprise organizations, including those that require deployments inside customer-controlled or on-premise environments.

Looking forward, Noah Labs is building deeper automated verification features so engineers can even more easily assess changes and trace the evidence behind every modification, while exploring how Gemini platform and Gemini models can further expand cloud deployment options alongside local installations. Isik adds, “with Google DeepMind’s Gemma 4 31B and Google Cloud, Noah Labs is advancing practical AI for legacy software modernization, helping engineering teams turn complex codebases into understandable, testable, and maintainable systems."

Noah Labs builds AI tools for mission-critical software and autonomous systems, helping teams modernize legacy code and monitor navigation reliability.

Industry: Technology

Location: United States

Products: Compute Engine, Gemma 4, Gemini, Google Cloud

Google Cloud