Decoding LLMs: Insights and Challenges on the Path to AGI
Like many of you, I closely follow the progress in AI and LLMs , particularly in the quest for AGI. In this article, I delve into the fascinating complexities and hidden mechanisms of large language models (LLMs), shedding light on their capabilities and limitations. LLMs, such as ChatGPT-4, represent a remarkable leap in AI, but understanding their true potential requires unraveling the intricate processes within their “black box.” I’ll explore how these models dynamically transform simple prompts into detailed responses, the curious phenomenon of hallucinations, and the critical role of context in their performance. Additionally, I will discuss my take on the future of AI, including the integration of specialized AIs into cohesive networks and the ongoing challenges of achieving true artificial general intelligence. This article is not heavy on technical details and was written for a broad audience. Read More »
How to Inject Custom Logic to Intercept and Modify API Requests and Responses Without Touching Frontend/Backend Code

When working with APIs, dealing with raw JSON for input and output can be cumbersome, especially when there’s a convenient UI designed specifically for that API. However, the UI doesn’t always keep up with changes in the service, and sometimes it’s useful to emulate changes in the service along with the UI to finalize decisions on what and how to modify.

Such a solution could allow you to replace actual API responses with tailored fake data, eliminating the need for altering real data or making code changes. It would accelerate slow requests by recording and replaying them, significantly reducing wait times for UI tweaks. Developers could simulate edge cases, such as handling 4XX or 5XX status codes or testing long strings, by modifying response bodies, status codes, or headers, or add a debug/verbose mode automatically to all calls to an API. Additionally, it would enable frontend development against unimplemented APIs by mocking responses, facilitating progress even when backend endpoints are incomplete. For integration with third-party platforms that only function on production domains, such a solution could reroute requests, allowing safe testing on production without risk. Furthermore, it would support debugging by adding artificial delays and simulating network errors, ensuring robust handling of edge cases. In essence, mocking servers on production in your browser would streamline development, enhance testing, and ensure a more resilient application.

As the “victim,” I chose the SAP Commerce demo store (to whose code I don’t have access to), specifically its search functionality. The goal was to inject my custom code between its backend and frontend, and as a demonstration of capabilities, automatically apply certain facets for specific queries. As a result, for the query “cheap cameras,” I will actually see cheap cameras and not tripods, for example.

Read More »
Tracking and Visualizing Real-Time Operational Metrics in E-commerce

In one of my projects, it was crucial for me to monitor data volumes and some specific e-commerce system parameters and alert the technical team if there were any abrupt or unexpected changes in them. These fluctuations occasionally (though not always) signaled potential existing issues or an increased likelihood of future problems.

I am focusing on operational parameters (orders, shopping carts, customers, logins, product data quality), not technical (page load time, timeouts).

Due to the lack of built-in solutions in SAP Commerce for this function, I had to design and implement a custom solution from the ground up. This article delves into the specifics of that development process and details of the architecture.

Read More »
Saturday Evening Post Archives: Engineering a Smart Search Solution

In one of my recent personal projects, I ventured into the vast archives of The Saturday Evening Post, a magazine that has captured American life since 1821. What started as a nostalgic dive into the past tu,/rned into a technical quest to make over a century of magazine issues searchable and accessible online.

Utilizing a blend of OCR with Tesseract and search mechanisms based on Apache Solr, OpenAI embeddings, FAISS, ML Re-ranking, I managed to create a searchable database of the magazine’s rich content from 1900 till our days. From Norman Rockwell’s iconic illustrations to pivotal historical articles, every page was digitized and ready to be explored.

This project engagingly occupied several weekends and further deepened my expertise in search technology. Dive into the article to see how vector search and traditional keyword search can be used to bring the past into the digital age.

Read More »

Last night, a spark of inspiration on my evening drive turned into a monumental achievement by the early hours: I successfully migrated 6,000 posts from Facebook to my brand-new Wordpress blog and later translated them to English and published at So now I have two blogs, in Russian and English, having the same content as my Facebook (I don’t post anything on Facebook that isn’t meant for others to see, and I don’t write or say anything that I wouldn’t say publicly, so from a privacy perspective, everything is okay. Also, I don’t transfer comments to the posts, only the posts themselves.) This isn’t just about launching another blog; it’s about preserving years of shared memories and insights in a more accessible and enduring format.

I tackled the challenge of exporting and translating a vast archive of content, making it available not only in the original Russian on but now also in English for a global audience. This project was driven by my frustration with social media’s fleeting nature and the limitations in searching tagged posts.

Curious about how I managed to automate the complex process of syncing thousands of posts across languages and platforms? I’ll be sharing the behind-the-scenes story of the tools and technologies that made this possible.

Read More »

Navigating the vast seas of product searches in B2B and B2C marketplaces can be quite a headache, especially with loads of data from various suppliers. It’s tough for marketplaces to tweak supplier info without spending a fortune, yet when searches go awry, customers tend to blame the marketplace, not the suppliers. My new article suggests a slick new way to cut through the clutter—think product families, not just standalone products. This isn’t just about tidying up; it’s about making the search experience smoother and more intuitive for everyone involved.

Let’s dive into why the old way of doing things—splitting catalogs into neat categories and individual products—just doesn’t cut it for big, diverse inventories. Take electronics and clothes: sorting shirts by color and size is straightforward, but try categorizing gadgets that way and you’re in for a real challenge. The game-changer here could be switching to a product families model, which groups related products together under one big family umbrella. This approach keeps things organized and makes shopping easier, which is exactly what we need for a smoother, more customer-friendly search process in our digital storefronts.

The Part I of this topic is about Product Families as an extension of the Variant Products concept.

Read More »

An incremental load is a form of ETL (Extract, Transform, Load) process that involves transferring only the data that has been altered since the last loading session. This approach is both more efficient and quicker than performing a full load because it minimizes the volume of data that needs to be moved and processed. 

However, implementing an incremental load necessitates mechanisms for recognizing and monitoring modifications in the source data, using methods like timestamps, flags, or versioning. Additionally, it must effectively manage issues like conflicts and duplications, for instance, when a single record undergoes multiple updates.

In this article I am focusing on my experience in implementing the optimized data loader integrated into the target system which was SAP Commerce. Simply put, it handles duplicates and minimizes the number of unnecessary database writes. 

Read More »

Imagine a world where an AI system not only listens but observes, constantly collecting data unless explicitly told not to. This “exoskeleton for the brain” is not just a tool but a transformative extension of human capability, enhancing efficiency and decision-making in everyday tasks. For some, this raises alarms about privacy and the potential for misuse in a society already wary of data overreach. Yet, the real question lingers: how will such capabilities shape the way we live, work, and interact not just with technology but with each other?

As we stand on the brink of making these integrations a reality, it’s important to consider not only the current implications but also the generational shift that might come with it. The youth of today, growing up with AI as a norm rather than an addition, might view this integration as indispensable—melding their cognitive functions with artificial intelligence to an extent previously unimagined. This narrative isn’t just about the technology we create but also about the human experiences and societal structures that will evolve with it. Dive into the full discussion to explore the profound shifts looming on the horizon.

Read More »

In this article, I want to share an interesting experience. It has nothing to do with SAP Commerce and might even be unrelated to eCommerce, but it could be useful for Solution Architects when tackling a task similar to mine. And the task was interesting.

It was a huge Excel file for generating estimates for B2B customers. A manager entered a bunch of parameters according to customers’ needs and get a printable quote. This Excel spreadsheet was really old, already about ten years old, and all these ten years, different people have been adjusting it, correcting it, adding formulas and constants. Basically, almost nobody knew how it worked in all details. All this evolved into a dozen tabs, each filled with numerous formulas. And these formulas refer to similar formulas from the adjacent tabs. It has almost a thousand formulas, mostly complex. How to untangle this knot?

There was no documentation, but even if there had been, where is the guarantee that the documentation fully matches the Excel file? After all, if it doesn’t match, finding the root causes of the discrepancies will be very difficult. Since the Excel file is the source of truth, it basically is the documentation.

These are the kinds of problems I love. In the article, you’ll find the solution and, believe me, it’s very non-trivial.

Read More »
Search Analytics: Testing and Monitoring eCommerce Search

In this article, I want to talk about my experience with how to make a website or online store search more responsive to business requirements and user expectations.

It’s common to give too much attention to technical issues, especially now, with the advent of machine learning/AI in this field. But more often than not, the deepest problems are on the surface and their solution is much simpler than a functional refinement.

One of the main challenges in managing e-сommerce websites is ensuring the product search feature works effectively. Imagine this: you notice an issue with how products are being found through search, and you decide to fix it by tweaking some settings. However, altering these settings could unexpectedly impact your business. When we talk about something being ‘broken’, it’s not just about error messages popping up. Rather, it could mean that customers are having a harder time finding what they want, which might lead to fewer sales over time. This is a subtle effect that can spread across your entire platform. But how do you pinpoint the exact cause of such a decline? Testing and monitoring changes in an ecommerce environment are complex because even small modifications can have wide-ranging effects on customer behavior and, consequently, sales performance.

To monitor changes, it’s essential to understand what to collect and how to extract data from the gathered information. Specifically, for e-commerce search, there are no good ready-made monitoring tools available, but there are various DIY kits that address some needs. Such tools suggest collecting and analyzing search results for reference queries and visualizing changes over time. However, this approach falls short in many respects. For instance, it only covers the most frequent queries.

When it comes to modernizing site search, there’s a big question about whether this is done from scratch or if it’s possible to use accumulated data from existing sessions. Developing from scratch is always a risk—because any assumptions you have about how users will search for content may be wrong. Therefore, regardless of how good the first version of the search turned out to be, it will need to be revised and reworked many times.

Today, I want to share the approach that I have used in several projects. I call it Search Analytics. Read More »