Pebble - Blog by Stuart HirstPebble - Blog by Stuart Hirsthttps://www.pebbletek.ai/blogs/author/stuart-hirstThu, 21 Aug 2025 18:58:37 +1000http://zoho.com/sites/<![CDATA[Agentic Observability]]>https://www.pebbletek.ai/blogs/post/agentic-observabilityExplore the transformative potential of agentic AI and the critical role of observability. As AI autonomously plans and executes tasks, transparency is essential. PebbleAI.app provides dashboards, real-time monitoring, and collaboration tools, enhancing security, performance, and compliance.]]>

The Rise of Agentic AI and the Need for Observability


​

​

​

​

​


There’s a new buzzword atop Gartner’s tech trend list: Agentic AI. This refers to AI systems (often powered by large language models) that don’t just respond to a single prompt, but can autonomously plan and execute tasks – in other words, AI with a degree of agency. From customer service bots that juggle multiple queries to AI “assistants” that can browse the web and trigger workflows, these agents are quickly moving from sci-fi to business reality. Gartner projects that by 2028, agentic AI will be making 15% of day-to-day work decisions and that one-third of enterprise software will include agentic AI components (up from less than 1% today).


As exciting as this is, it also poses a burning question: How do we monitor and trust these autonomous agents? In traditional software, we use logging, monitoring, and APM tools to know what’s going on under the hood. We need the same (if not more) for AI agents that learn and act in semi-unpredictable ways. This is where agentic observability comes in – a new discipline focused on gaining visibility into the inner workings of AI agents.

Observability in this context means capturing the full story of an AI agent’s operation: the prompts it received, the chain of thoughts or reasoning it went through, every call it made to an external tool or API, and the outputs it produced. Why is this critical? Because without it, deploying an AI agent is like employing a staff member you can’t supervise at all. You wouldn’t run a business department in total darkness, and similarly, you shouldn’t deploy autonomous AI without the ability to observe and audit its decisions.


Current Solutions and Their Shortcomings

Recognising this need, the market has seen an emergence of tools aiming to monitor AI agents. Some are open-source projects born out of the LLM developer community, while others are commercial platforms from startups and incumbents. For example, LangSmith (from LangChain) and Langfuse offer trace logging of LLM calls; tools like Helicone act as proxies to capture prompt/response data; and Arize/Phoenix and WhyLabs are extending traditional ML monitoring into the generative AI realm​. This flurry of innovation is great to see – it shows that as an industry we’re taking AI oversight seriously.

However, most of these solutions address only parts of the problem. A lot of developer-focused tools can log each API call or each prompt, but they struggle to present the higher-level picture – especially when an agent has to string together dozens of steps or collaborate with other agents. Imagine a customer support AI agent that, to resolve one query, might consult a product database, draft an answer, double-check via another AI, and then respond. A human debugging that session needs to see the whole sequence with timestamps and decisions. Few tools today provide that out-of-the-box.

Another issue is fragmentation. One tool might tell you how many times the AI called the language model (and how much that’s costing you), while another might evaluate whether the AI’s answer was factually correct. You might resort to custom scripts or manual analysis to connect the dots – hardly a sustainable strategy. This fragmentation means higher overhead to achieve observability, and potential blind spots if some piece isn’t captured.

Crucially, many existing solutions weren’t built with enterprise governance in mind. They were made by and for AI developers, which is fantastic for early-stage experimentation, but enterprises have stricter demands. Logging data may need to be kept private or on-premises. Different teams might want role-based access (e.g., only data scientists can view raw prompts with customer data, whereas business owners just see aggregate metrics). Compliance logging – say, keeping records to comply with the upcoming EU AI Act – might not be considered in a hobby project that focuses just on debugging model prompts.

In short, while we do have tools signalling the start of “AI observability” as a practice, there’s a gap in integrated, enterprise-ready solutions that cover the full lifecycle of agentic AI. Businesses adopting AI agents today often express frustration that they lack a “single source of truth” for understanding their AI’s behavior. They might be cobbling together an internal dashboard or exporting logs to Excel – approaches that won’t scale when they go from one pilot agent to a fleet of fifty automating various processes.
Our Vision: The Central Observatory for All Your AI Agents

Holistic View

Pebble AI doesn’t just show you API calls or chat transcripts in isolation. It reconstructs the full narrative of each agent’s task. If your AI sales assistant goes through 10 steps to generate a lead, you’ll see all 10 in order – along with timing, cost, and outcome at each step. This holistic trace is presented in an easy, visual timeline UI, so you can literally replay an agent’s thought process. Think of it like debugging a complex program – but the “code” is the agent’s decisions and the “debugger” is Pebble AI, stepping through each function call (be it an LLM call or a database query).

Integrated Governance Features

We know observability data isn’t just for engineers – it’s for compliance officers, risk managers, and business leaders. Pebble AI comes with audit trail exports and compliance dashboards out-of-the-box. Need to comply with that EU AI Act logging requirement? Your Pebble console can generate a report of all agent decisions, complete with timestamps and annotations about who/what reviewed them, in a format ready for auditors. We also allow setting policies in the platform – for instance, “if the agent’s confidence is below X, require human review” – and Pebble will help enforce those by catching low-confidence outputs and alerting a human operator. By baking governance in, we remove the need for separate manual processes to supervise the AI; oversight is part of the system’s DNA.

Multi-Agent Aware

Organisations are starting to deploy not just one agent in a silo, but suites of agents that might collaborate. Pebble AI is built to handle that reality. You can monitor not only individual agents, but also the interplay between them. For example, if you have one AI agent summarizing data and another agent taking that summary to draft a report, Pebble will show the handoff. We support tagging and grouping of agents, so you can view metrics for a single agent or aggregate for a team of agents working on a process. This is agentic observability at the system level, not just the component level.

Intelligent Insights, Not Just Data

Raw logs are not enough. Pebble AI is incorporating analytics and AI to help interpret the observability data. Our platform can automatically flag unusual patterns (using anomaly detection models) and even score the quality of agent outputs using built-in evaluators. For example, Pebble might highlight that “This week, Agent A’s average task completion time doubled compared to last week” or “Agent B has produced answers with a much more negative tone than usual – possible issue?”. These insights are surfaced in a dashboard and can be sent as alerts. The idea is to offload the burden of analysis – Pebble doesn’t just collect data, it helps you make sense of it at a glance. Ultimately, this means you spend less time sifting through logs and more time optimizing and strategizing your AI deployments.

Flexibility and Integration

Pebble AI is designed API-first. We know each company might use a different stack – you might have some Python agents using LangChain, some Node.js services calling the OpenAI API directly, maybe an off-the-shelf chatbot platform – and ripping those out is not feasible. So instead, Pebble AI provides connectors and SDKs to plug into any environment. We want to meet you where you are: whether it’s via a simple REST endpoint where you send us logs, or a direct integration with your cloud provider’s AI services, or a drop-in library that instruments calls. Similarly, we make it easy to pull data out: want to push an alert to Slack or your SIEM when an agent makes a critical error? Or query our system from your BI tool to see quarterly trends? All doable. Pebble AI acts as a hub, not a closed box. This stands in contrast to some solutions that lock you into their interface or require all-or-nothing adoption.

Enabling Sustainable AI Automation

When agentic AI is deployed with proper observability and control, it becomes a reliable engine for automation. We move past the fear of “AI behaving badly” because we have the tools to catch and correct any misbehavior. This opens the door for businesses to automate more and innovate faster with AI.


Picture a future (not far off) where every company has dozens of specialized AI agents: researchers, assistants, analysts, customer reps. Managing that could be chaos – unless you have a way to see and manage all agents centrally. Pebble AI aspires to be that command center that makes large-scale AI agent deployment not only possible, but smooth and sustainable.


At Pebble AI, we believe transparency is empowerment. By shedding light on AI agents’ inner workings, we empower businesses to embrace these agents confidently. Instead of treating AI like a mysterious black box or a quirky intern that sometimes “figures things out somehow,” companies can treat AI agents as fully accountable team members – whose work is tracked, quality-checked, and aligned with the company’s goals and values.


The importance of agentic observability cannot be overstated. It’s how we ensure AI continues to serve us, not stray into unintended territory. It’s how we scale one successful pilot into a thousand agents creating value across an enterprise. And it’s how we build trust – with users, with regulators, and with ourselves that we are in control of our creations.


Pebble AI is proud to be at the forefront of this movement. We’re not just building a product; we’re helping shape the practices and standards that will define responsible AI automation in the years to come. If you’re on the journey of adopting AI agents, we invite you to partner with us. Together, we can unlock incredible productivity while staying true to the principles of transparency and accountability.


Ready to shine a light on your AI Agents?

Join the waitlist for  Pebble AI’s platform and see how agentic observability can take your AI initiatives to the next level. Your AI agents don’t have to operate in the dark – and neither should you.
Get PebbleAI
]]>Fri, 21 Mar 2025 11:21:55 +1100<![CDATA[Real-Time AI Voice, Wow]]>https://www.pebbletek.ai/blogs/post/OpenAI-Real-time-Voice Imagine having a virtual assistant with the voice of your choosing and specialist knowledge that you integrate. This technology could allow ]]>

OpenAI Dev Day 2024: Advanced Voice Mode

OpenAI's Revolutionary Voice Technology: 
A Game-Changer for AI Interactions

In a groundbreaking development, OpenAI has recently unveiled its latest advancement in AI technology: highly realistic voice synthesis capabilities. This innovation is set to transform the way we interact with AI systems and opens up a world of exciting possibilities across various industries.


Just to be clear; this is not just another text to speech service or a voice cloner, this new capability enables OpenAI models such as ChatGPT and GTP4o to hold a real-time conversation with the user. You can try it for yourself in the OpenAI IOS App.


Understanding OpenAI's Voice Technology

OpenAI's new voice technology leverages advanced machine learning algorithms to generate human-like speech that is nearly indistinguishable from real human voices. The system can clone voices with remarkable accuracy, mimic emotions and intonations, and even speak in multiple languages all in real-time.

Potential Use Cases

Personalised Virtual Assistants

Imagine having a virtual assistant with the voice of your choosing and specialist knowledge that you integrate. This technology could allow users to customise their AI assistants with familiar voices, making interactions more comfortable and engaging whilst enabling simplified access to specialist knowledge.

Language Learning & Translation 

Language learners could greatly benefit from AI-generated speech in various accents and dialects, providing a more immersive and authentic learning experience. This technology could offer personalised pronunciation feedback, simulate real-world conversations with native speakers, and adapt to the learner's proficiency level. Moreover, it could facilitate real-time translation, allowing learners to practice their target language while instantly receiving translations in their native tongue.

For translation purposes, this AI voice technology could revolutionise cross-language communication. It could provide instant, natural-sounding translations in various languages, complete with appropriate accents and cultural nuances. This would not only aid language learners but also break down language barriers in international business, tourism, and global collaboration efforts.

Accessibility Solutions 

For individuals with visual impairments or reading difficulties, this AI voice technology could be truly transformative. By providing more natural and expressive text-to-speech and vision-to-speech solutions, it can significantly enhance their digital experiences and daily lives. 

These advancements could lead to greater independence, improved access to information, and enhanced quality of life for millions of people with visual or reading impairments worldwide.

The ability to process vision and interact in real-time could be a great update for our VisGuide project. Follow as we update VisGuide and create VisGuide Real-Time.

Customer Service

Businesses could implement more human-like AI voice agents for customer support, revolutionising the customer service industry. This technology has the potential to significantly improve customer satisfaction while reducing operational costs for businesses.

We are working an example of this to demonstrate the value of connecting this capability with the real world.

Care Companions

AI-powered care companions could provide 24/7 emotional support and companionship for elderly or isolated individuals. Using OpenAI's voice technology, these companions could adopt familiar voices, engage in natural conversations, and even reminisce about shared memories. They could also monitor health patterns, provide medication reminders, and alert caregivers in case of emergencies. This technology has the potential to significantly improve the quality of life for many, offering comfort and assistance while maintaining a sense of independence.

Ethical Considerations

While the potential applications are exciting, it's crucial to address the ethical implications of this technology. Issues such as voice cloning consent, potential misuse for disinformation, and privacy concerns need to be carefully considered and regulated.

The Future of AI Voice Technology

As this technology continues to evolve, we can expect to see even more innovative applications. From more natural human-AI interactions to revolutionary changes in how we consume and create content, OpenAI's voice capabilities are set to play a significant role in shaping our digital future.

In conclusion, OpenAI's new voice technology represents a significant leap forward in AI capabilities. As we explore its potential within Pebble, it's important to balance innovation with responsible development and use. The future of AI-human interaction is here, and it sounds more human than ever.
]]>Sun, 01 Sep 2024 09:00:00 +1000<![CDATA[VisGuide - A Generative AI Project for Visual Assistance]]>https://www.pebbletek.ai/blogs/post/visguide-a-generative-ai-project-for-visual-assistanceVisGuide, a Pebble-developed AI project, aids visually impaired users by providing verbal descriptions of surroundings. Using OpenAI vision models and ElevenLabs voices, it offers navigation assistance and scene descriptions, showcasing AI's potential to enhance accessibility and independence.]]>

VisGuide - Context & Part 1

A while ago we created a project to explore the capabilities of AI and Vision to see if we could create a cost effective solution, with commodity hardware, that would have a meaningful impact and improve someones quality of life using AI. 

We created a solution called VisGuide that provided the user with a verbal description of their surroundings and helped to both guide and keep them safe, empowering their independence and participate in experiences that they had previously been excluded from. 


As you will see below, we managed to create a pretty cool solution and provided the code as an Open Source project. With the recent release of Open AI's Real-Time Voice API we thought it was time that we revisited this project and refreshed it based on the new capabilities. Part 1 (original VisGuide) is below and we are actively working on Part 2, watch this space......



Motivation & Inspiration

GenAI is constantly in the news and the rate of innovation is like nothing we have seen before. There is no doubt that it will change lots of things but somethings are more obvious than others. It’s great at building websites, writing blog posts, creating pictures and much more but what intrigues me is how do these capabilities help to solve real issues rather than just being “cool”.

This project was inspired by another project called “Narrator” that has David Attenborough narrate what he sees from your webcam. Its cool little project that chains OpenAI and ElevenLabs to do something fun. Check out Narrator here https://github.com/cbh123/narrator

As someone who only dabbles in coding for simple tasks, perhaps once a year, my familiarity with Python could best be described as basic. Every time I use it I effectively have to relearn it. And so, the scene is set; can I use GenAI to both simplify creating the code and have it power something useful. This led to “What if it was mobile?”, “Could it help guide someone and help avoid danger?”

Project Scope

The aim of VisGuide is straightforward: to provide visual assistance to the partially sighted or blind using commodity hardware and GenAI. This initiative explores the potential of Generative AI to serve a practical, impactful purpose, reflecting a broader trend towards technology solutions that enhance quality of life for individuals with disabilities.



VisGuide Foundations

VisGuide’s hardware setup is simple, utilising a Raspberry Pi Zero with a camera and a power bank for mobility. Connectivity is achieved through a mobile phone hotspot, enabling the device to operate in diverse environments. The choice of hardware underscores the project’s commitment to accessibility and affordability.

The software, written in Python, is the core of VisGuide. It incorporates:
  • Generative AI: Utilising OpenAI’s vision models, VisGuide interprets visual data to provide verbal descriptions of the user’s surroundings.
  • ElevenLabs: The platform is used for voice synthesis, ensuring the verbal feedback is both clear, natural and that the text to speech process is as fast as possible.
  • ChatGPT & Github CoPilot: The development process leveraged a combination of ChatGPT and Github CoPilot for code development. I found that both worked well but mainly used ChatGPT as the user experience was simple and when it didn’t do what I needed, I just changed my prompt to be more specific.


Open Source Code

You can run it on your Mac to test it and it runs on a Raspberry Pi Zero.
It is not yet optimised for speed and runs a little slow on the Raspberry Pi but works much quicker on the Mac.

GitHub

The VisGuide project is hosted on GitHub and Open Source so you can download it and get it working locally to experiment with the combination of AI and Vision:



Heres an example of the “guide” narration:

00:00
  • narration_example
    00:00

Modes & Use's

VisGuide can be run in either continuous or on demand mode. On demand mode is where the user presses a button to capture the scene and have it narrated and continuous mode is where it runs in a loop every 5 seconds.

VisGuide has two primary use’s, the first is as a visual assistant or “guide” and the second is “tourist” mode where it describes the scene with colourful language to create a mental picture for the user.

  • Guide Mode
    This is intended to provide a brief and succinct description of the scene to help the user better navigate and avoid obstacles and risk.

  • Tourist Mode
    This is where the user can participate in “sight seeing” and have a colourful and artistic narration of the general scene. Why shouldn’t partially sighted and or blind people be able to participate in “sight seeing”. This, combined with the users other senses, helps users to appreciate what their sighted peers can experience.


Impact & Potential

VisGuide is a working example of digital tools and services helping with a physical problem and explores how it can make a real difference. It explores offering increased independence for visually impaired users and showcases the potential of Generative AI to address and solve real-world challenges.


Whats Next?

  • Interactive Uses
    It would be ideal to enable the user to “query” the scene. For example the user might ask “Where is the road crossing” or “Please guide me to the door”. This is where user input updates or augments the prompt to influence the narration response.
  • Integrated Hardware
    Wearing a Raspberry Pi around your neck with a battery pack is hardly ideal but great for development. Ideally, VisGuide would work with something like Meta’s Smart Glasses which have the camera and audio output built in and would provide a unobtrusive user experience.
  • Local Processing
    VisGuide currently requires internet access to leverage the SaaS LLM services. Mobile phones are incredibly powerful compute devices these days and so it would be ideal to have all processing performed locally and remove the dependency on connectivity and the internet however this would significantly increase the complexity.

Conclusion

VisGuide’s development journey illustrates the power of combining simple tools with cutting-edge AI to create solutions that have a real-world impact.

As the project moves forward, it’s intended to be thought provoking and demonstrate the profound capabilities of technology when applied with purpose and vision.

This project is more than just a technological achievement; it’s a pathway to greater accessibility and independence for those it aims to serve.

Part 2 is in development and we hope to share it soon.
]]>Mon, 01 Jan 2024 08:00:00 +1100