SevenPico Insights

Quarter 1 2025: AI Insights

Written by Brandon Oddo | Apr 7, 2025 2:14:41 PM

This last quarter has been heavily focused on AI. Here are highlights from this past quarter of newsletter emails that Eric and I wrote:

 

SevenPico's Bespoke GenAI Agents Supercharge Domain Knowledge Retrieval

Generic AI can't solve unique business problems. Despite their capabilities, ChatGPT's web search and Microsoft 365 Copilot's document access lack the fine-tuned controls and contextual understanding your organization requires. SevenPico has bridged this critical gap with a custom Slack-integrated AI assistant built on AWS Bedrock Agents and Knowledge Base, featuring precision-engineered retrieval systems specifically calibrated to your domain expertise and workflow requirements.

The bespoke Knowledge Base and Agent configurations allow the AI to provide robust citations for its answer including links, tags or categories, and even relevance scores. Our solution enables team members to ask natural questions like "Where in the codebase does user authentication live?" and receive accurate, contextual answers instantly. 

A custom knowledge assistant is ideal when you need:

  • More than general web search: Access proprietary information, especially confidential info requiring enhanced security, and knowledge that stays current
  • High accuracy: Verified responses with reduced hallucinations, thanks to granular guardrails
  • Thought Process Tracing: Enhance decision quality by customizing how information is organized and tagged, allowing search results to be prioritized for specific teams, filtered with precision, and supported by direct citations to trusted source materials

Imagine the possibilities:

  • Put your organization's collective expertise at everyone's fingertips, eliminating fruitless searches for that critical document that "I know is here somewhere"
  • Deploy Customer Support chatbots that consider customer-specific context, can process refunds, create support tickets, and make wiser pre-escalation decisions
  • Leverage event-driven architectures to enable AI that reacts instantly to system events, providing immediate insights for faster, better decision-making

 

SevenPico Gains 2.5x Development Velocity with AI Coding Agents in an Established Codebase

Conventional wisdom in Q1 2025 states GenAI coding agents only work on brand new codebases. Up against the size and complexity of established codebases, they fall flat. With Prior Proper Planning, SevenPico found you can prevent that Piss-Poor Performance everybody is talking about.

We've successfully coded production-ready revisions in our client's large codebase using GenAI agents. The experiment: revise all logging in the codebase to improve observability, performance and cost efficiency. Our secret? Adapt the same process SevenPico has been using with offshore developers for years:

  1. Craft a comprehensive Design Document
  2. Break it down into small, well-defined, testable units of work
  3. Write tests to verify each unit's code works
  4. Write code to pass the tests
  5. Code review everything skeptically
  6. Confirm each unit's behavior with a quality assurance engineer

The GenAI code editor, Cline, uses different agents for planning and coding activities: In Plan mode, the agent accelerates many design tasks, reading our project brief and relevant code in the codebase while building its long-term Memory Bank. It quickly elaborates on our brief to create a comprehensive Design Document with per-ticket specifications. With these inputs, Cline's Act mode then iterates to smoothly implement the small, well-specified tickets without hallucination, under minor supervision. Since our tickets always involve test-driven development, the agent must execute automated tests and iterate until it can factually assert all works as intended. 

In our early experiments we averaged 2.5x faster velocity than the AI time estimate per human-developed ticket. We are targeting 4x velocity with a bit more practice and believe even faster is possible if we parallelize tickets across multiple agents. With GenAI, even for established codebases, huge velocity boosts are possible. SevenPico can help you get there.

 

Building a Wiser AI Agent Using a Retrieval Augmented Generation (RAG) Knowledge Base

A marketing agency brought us an interesting use case:
 
How can I leverage AI to become an expert about my prospective client's existing marketing presence before my first meeting with them?
At SevenPico, we recently ran an experiment comparing two AI-powered approaches for rapid client research:
  1. ChatGPT 4o with web search
  2. Custom AI Agent with a RAG (Retrieval Augmented Generation) Knowledge Base (AWS Bedrock with Claude 3.5 Sonnet v2)
Why RAG?
A RAG Knowledge Base lets an AI agent reference your unique expertise without confusion—pulling in up-to-date and specialized documents that go beyond surface-level web information.
 
Key Takeaway
Both solutions delivered quick, high-level competitive analysis. ChatGPT benefits from low setup time and cost and provides quick, general insights based on public websites. However, the Knowledge Base Agent leverages richer, more specialized data sources (social media, SEO reports, etc.), tapping proprietary data in real time for deeper, more relevant results. The Knowledge Base Agent's advantage grows the more it is fed hard-to-web-search information in its knowledge base. Potentially both agent types could even collaborate, with one agent-with-search placing results into the knowledge base to be cited by the other.

 

SevenPico Cut Developer Cycle Times by 1 Day Using AI Code Review

Whether you buy or build technology, software development always takes longer than you want it to. For example, while code reviews are essential, traditional peer review processes can slow progress due to delays in handing off work and during the review itself. SevenPico automated code reviews using GenAI with CodeRabbit.

By switching from peer code reviews to AI-driven reviews, we eliminated more than one calendar day of cycle time per ticket, as the developer can immediately engage with the AI as soon as the implementation is ready for review. We also saved roughly an hour of active reviewer time per ticket—likely even more for complex changes. We also refine the code review instructions for the AI as we go, continually improving code review quality for the entire organization. All together, we saved a lot of time without noticeable increases in defects or rework.

How SevenPico AI Recovered Institutional Knowledge After Losing Key Expert

After a key employee leaves, how do you keep going if they've taken important institutional knowledge with them? 

We helped one of our partners operate their technology after a key employee left. Staff reported that something customer-facing in their application had broken: Without solid documentation or a knowledge transfer, how could we quickly solve the problem?

We used Cursor, a powerful AI code editor that indexes all source code, enabling you to chat about its contents. By leveraging AI, your application code can serve as living documentation for your company, holding valuable insights—especially when developers consistently comment their code. We described the problematic behaviors we were observing and requested a codebase-wide semantic search using "@codebase". This powerful AI technique, called Retrieval Augmented Generation (RAG), helped us tap into the code and find the root cause in minutes instead of the hours it would have taken a human to review it. We discovered that the root cause of the problem was a misconfigured value on the administrative side of the system. After making that simple update, the issue was resolved immediately.

 

SevenPico Reduced Waste in Only 15 Minutes of AI Consulting

In our experience, there are almost always cost savings opportunities in your business's information systems. For example, looking through your application error logs can be about as interesting as filing your taxes: Both can be tedious, dreadful tasks, but both can also lead to unexpected financial rewards. We experimented with AI analysis of system activity logs and found inefficiencies that helped us cut unnecessary costs. 

We analyzed AWS CloudWatch error logs for a customer’s application and fed them to Cursor. This led us to an important discovery—the system was making unnecessary API calls, which were driving up costs. After reviewing the AI's suggestions, we added validation logic to requests before they would be sent, ensuring only meaningful data would be processed. This simple fix eliminated a high volume of wasteful API calls and so immediately reduced costs. This experiment is so new that we don't yet have a month's worth of total cost savings from this example, but the impact is clear—less unnecessary compute, lower bills, and a more efficient system.

 

Ready to leverage your organization's knowledge, streamline development, and cut costs with AI that works specifically for you? 

SevenPico delivers secure, customized AI assistants built on your private network with AWS Bedrock Knowledge Base integration. From improved documentation to AI-powered code reviews, we're helping teams work smarter across industries.

Contact us today to turn your business data into actionable intelligence and watch your team's productivity soar.