Bex Tuychiev

📢 Introducing My Research Project: Cluster Ruler – Real-Time Adaptive Urdu Sign Language Clustering 🚀 Proud to share my final year research project: Cluster Ruler — a system designed to recognize and cluster Urdu/Pakistan Sign Language gestures in real time, without the need for manually labeled data! ❌📝 ✅ Key Highlights: • 🔍 No frame-by-frame labeling required — clustering is fully unsupervised • 🎥 Real-time gesture detection using webcam and MediaPipe • 🧠 Extracts temporal hand landmarks and features for adaptive clustering • 🔄 Built using a Simulated SVTFormer model for high-dimensional feature encoding • 📂 Dynamically creates, renames, and merges gesture clusters on the fly • 💻 Powered by Python (Flask), MediaPipe, HTML/JS — runs entirely offline This project contributes to making sign language technology more scalable and accessible by removing the bottleneck of annotation. Aimed at bridging communication gaps and empowering accessibility solutions. If you're into machine learning, computer vision, HCI, or accessibility innovation — let’s connect! #Research #ClusterRuler #SignLanguageRecognition #UnsupervisedLearning #MediaPipe #DeepLearning #FinalYearProject #ComputerVision #AdaptiveAI #GestureRecognition #Flask #PakistanSignLanguage #ML #HCI

13

1 Comment

So, Google went ahead and dropped a massive guide on how to prompt Large Language Models for FREE. Dives deep into many prompting techniques, best practices, examples and how to get predictable + reliable output (link below ↓) https://buff.ly/ozWxoLe

18

4 Comments

Sometimes you'll need better tools from top companies, even Google Years ago, when I was almost finishing up my undergrad during 2020, Fariz Rahman and me built a tool called TableQA, as we needed a tool that can query natural language on tabular data. Back then, our only alternative was to use Google's Tapas, which had issues with context length (will support only upto a limited number of rows, had memory issues etc) and we wanted a solution that could process infinite rows, connect to different SQL database sources, etc. And we built a tool, called TableQA, which performs natural language to SQL conversion and then applying the SQL on top of your data. I recieved so many emails, feature requests, and even several job offers just through this one tool. Fast forward to 2025... The emails still didn't stop. People still wanted a TableQA alternative, only difference is now we have LLMs, so they wanted LLM assisted query on natural language. This time I was already building my own company. So we converted such leads into our paying customers! At Datatune AI, we're commited to solving your issues with tabular data, and you get one of the best minds in the world at solving your problems!

32

🧠 𝗛𝗼𝘄 𝗱𝗼 𝗟𝗟𝗠𝘀 𝗹𝗲𝗮𝗿𝗻 𝘁𝗼 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗲 𝗰𝗼𝗵𝗲𝗿𝗲𝗻𝘁 𝘁𝗲𝘅𝘁 ? The key lies in causal (or masked) attention -> a mechanism that ensures the model only attends to past and present tokens, never the future. 🔍 During training, future words are masked out in the attention matrix. This enforces unidirectional context, for example, while predicting “you” in “Practice makes you better everyday”, the model only sees “Practice” and “makes”. It has no access to “better” or “everyday” yet. This constraint forces the model to learn language autoregressively, one token at a time ..... mimicking the way humans speak and write. This is how LLMs generate fluent, context-aware outputs without cheating. #AI #LLM #Transformers #CausalAttention #DeepLearning #NLP #AutoregressiveModels

14

Your First Client Will Teach You More About Your Work More Than Any Book or Course Ever Could! During my undergrad(Around 7 years ago) , I started freelancing on Fiverr(left freelancing about 4 years ago) as a data scientist. I had the skills Python, time series forecasting, and a stack of Kaggle projects but nothing prepared me for my first international client: X(Let's call it X), a boutique skincare brand in Barcelona. X had a problem: "We’re overstocking products that don’t sell and running out of the ones that do. Can you predict demand so we stop wasting money? I dove in, expecting a straightforward forecasting task. But their data was messy seasonal spikes, inconsistent sales logs, and no clear patterns. My first attempt? A classic ARIMA model. The results were… okay. But when I shared the forecast with the client, their response was: "This tells us what might happen, but not why or how to prepare for it." That’s when I realized: Prediction isn’t just about numbers it’s about action. I went deeper: Combined sales data with Google Trends to track interest in ingredients like "hyaluronic acid" or "vitamin C." Layered in weather data (yes, humidity affects skincare sales!). Built a hybrid model (XGBoost + Prophet) to predict demand and flag the "why" behind spikes. The result? X reduced overstock by 45% and cut storage costs in half. Even better, they started pre-launching products based on predicted trends - like a sunscreen line before a heatwave which boosted their revenue by 18% in six months. That project taught me: - A good prediction isn’t just accurate it’s actionable. - Clients don’t want models; they want decisions. Your first real-world project will teach you more than 100 tutorials.

7

Excited to share key insights from our latest research on Large Language Model #reasoning, co-authored with Yasiru Ratnayake! Our work, "𝗘𝗳𝗳𝗲𝗰𝘁𝘀 𝗼𝗳 𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 𝗼𝗻 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 𝗶𝗻 𝗶𝗻𝘀𝘁𝗮𝗻𝗰𝗲-𝗹𝗲𝘃𝗲𝗹 𝗦𝗲𝗹𝗳-𝗗𝗶𝘀𝗰𝗼𝘃𝗲𝗿", challenges the prevailing reliance on rigid, structured outputs for complex problem-solving. The quest for predictable #LLM reasoning often leads to structured formats like JSON. However, we introduced iSelf-Discover, an instance-level adaptation of the Self-Discover framework, to directly compare dynamically generated structured JSON reasoning with its unstructured, natural language counterpart. 𝗞𝗲𝘆 𝗙𝗶𝗻𝗱𝗶𝗻𝗴𝘀 𝗧𝗵𝗮𝘁 𝗦𝘂𝗿𝗽𝗿𝗶𝘀𝗲𝗱 𝗨𝘀: 1. 𝗨𝗻𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲𝗱 𝗟𝗲𝗮𝗱𝘀 𝘁𝗵𝗲 𝗪𝗮𝘆: Across diverse benchmarks, particularly the challenging MATH benchmark, unstructured plans achieved significant relative performance improvements of up to 𝟭𝟴.𝟵𝟬% over structured approaches. 2. 𝗭𝗲𝗿𝗼-𝗦𝗵𝗼𝘁 𝗣𝗼𝘄𝗲𝗿: Even more compelling, zero-shot unstructured iSelf-Discover variants outperformed their five-shot structured counterparts. This highlights the inherent strength of allowing LLMs more natural, free-form expression. 3. 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗶𝘀 𝗞𝗶𝗻𝗴 𝗳𝗼𝗿 𝗚𝗿𝗮𝗻𝘂𝗹𝗮𝗿𝗶𝘁𝘆: The optimal granularity of plan generation (instance vs. task level) isn't universal. It's context-dependent, varying with benchmark characteristics and the specific language model employed. 𝗪𝗵𝗮𝘁 𝗶𝘀 𝗶𝗦𝗘𝗟𝗙-𝗗𝗜𝗦𝗖𝗢𝗩𝗘𝗥? Our framework generates reasoning plans for each individual task instance, rather than a single task-level plan as in Self-Discover. It supports both structured (dynamic JSON) and unstructured (natural language) reasoning. The core operations are:  • 𝗦𝗘𝗟𝗘𝗖𝗧: Identifying relevant reasoning modules.  • 𝗔𝗗𝗔𝗣𝗧: Tailoring these modules to the specific task.  • 𝗥𝗘𝗔𝗦𝗢𝗡: Generating and then executing the plan. 𝗜𝗺𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻𝘀 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗙𝘂𝘁𝘂𝗿𝗲 𝗼𝗳 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀: These findings are not just theoretical; they have direct implications for designing advanced LLM agents. In fact, we're excited to be submitting our iSelf-Discover paper to the 𝗔𝗴𝗲𝗻𝘁𝗫 competition, a component of the "Advanced Large Language Model Agents" MOOC (https://lnkd.in/gj9KkfRp). Beyond AgentX, our work shows that while structured outputs are predictable, they can limit LLM reasoning on complex tasks. iSelf-Discover’s instance-level unstructured approach unlocks substantial gains, calling for a rethink of the structure-flexibility balance in AI. Future capable LLM agents may need more unstructured thought to leverage their full linguistic strengths. Our Pre-print will be available soon... We believe iSelf-Discover opens a promising avenue for exploring LLM reasoning capabilities. Looking forward to discussions and further explorations! #AI #ArtificialIntelligence #LargeLanguageModels #AIResearch #NLP #AIAgents #AgenticAI #LLMAgentsMOOC

61

2 Comments

A quiet warning for every nation and a loud one for Pakistan. Recent reports about foreign AI models such as Claude being used in sensitive military and strategic contexts linked to Venezuela and Iran should make us pause. AI is no longer just about chatbots, automation, or productivity hacks. It has crossed into the domain of national security, intelligence, and strategic decisionmaking. Here’s the uncomfortable truth: When a country relies on foreign AI models, it also relies on foreign priorities, foreign policies, and foreign control points whether visible or not. For Pakistan, this is an alarming signal. If we continue to consume AI without building it, we risk: • Strategic dependency • Loss of technological sovereignty • Limited control over critical systems • Long-term national security vulnerabilities History shows that nations which fail to own critical technologies eventually pay the price economically and strategically. AI sovereignty is no longer optional. It is not a “future goal.” It is a present necessity. The countries that build their own AI today will shape the balance of power tomorrow. The rest will simply adapt to decisions made elsewhere. The question is no longer if Pakistan should invest in indigenous AI, it’s how fast we can afford not to.

16

4 Comments

My understanding of AI agents so far: AI agents are dynamic loops of observe-think-act, where the model interprets context, plans next steps, invokes tools (function), updates its state, and repeats until the goal’s met. In other words, you prompt the LLM to decide which action or tool call to take at each step based on the evolving conversation or data. Good example is a sales agent that knows when to handoff the call to a human. ChatGPT when it understands that it should use search or/and fetch content from company knowledge base, combine both and show some visualization.

34

6 Comments

ML Lab | Day 8 (18th Feb 2026) - Data Collection with Web Scraping In last ML lab, we shifted focus from data analysis to data collection, an essential step before any ML model is built. We explored web scraping using: requests (GET & POST methods) BeautifulSoup for parsing HTML Extracting data using .find() and .find_all() Targeting specific elements like tables and rows Key takeaway: Before cleaning, visualising, or modelling data, you need to know how to collect it. Web scraping helps transform web content into structured datasets ready for analysis. Slow pace. #MachineLearning #WebScraping #DataCollection #Python #BeautifulSoup #MLLab #ComputerScience

25

2 Comments