My Activities

London Meetup Summary: Insights on Control Systems, LLMs, and AI Tools

At a recent London meetup, Serge Kozlov (Conundrum) explored deploying optimal control systems in factories. Using dynamic controllers with model feedback, he highlighted balancing material flow to prevent overload while maximizing yield. Model predictive control, trained via system identification, enables real-time parameter adjustments. The solution, requiring low-latency on-prem deployment, uses Kafka, ClickHouse, Kubernetes, and Python SDKs for 24/7 monitoring—proven effective in client projects.

Victor Zommers demonstrated visualizing LLM embeddings with T-SNE (Flask, Three.js, MongoDB), focusing on keeping LLMs relevant post-training. His pipeline integrates real-time macroeconomic data for traders, ensuring timely insights in fast-paced markets.

Lightning Talks:

• Tambe Tabitha Achere showcased automating GitHub workflows via Python (PyPI package), simplifying issue creation/management with tokens.
• Kevin Vegda critiqued AI framework pitfalls and built an AI interviewer with LangGraph, stressing scoping, production readiness, and tool selection.

A blend of cutting-edge tech and practical solutions—ideal for developers and engineers optimizing industrial systems, LLMs, or workflows!

Series of talks by Dojo employees and Google Startup Program manager on how to include diversity in product development and career tips for miniorities in tech.

Series of talks and workshop on Generative AI with Gemini at the center, third party cookies and flutter for mobile development.

Workshop on Agent Builder a no code LLM framework from google to orchestrate Agent workflows of LLM.

Series of talks on Generative AI concepts, with a talk from Nishi Ajmera (Lead Engineer @ Publicis) on Enhancing similarity search systems using RAG.

Spotting AMD Early: Our AI Innovation at NHS Hack Day

Age-related macular degeneration (AMD) affects 190 million globally, driving urgent need for scalable solutions. At NHS Hack Day, our team presented Deep Ocular AI—a tool leveraging a UNet-based architecture to analyze OCT retinal scans and tabular data, improving AMD detection and segmentation.

Current diagnostics rely on subjective, time-consuming manual analysis. Our model addresses this by combining multimodal inputs (cross-vendor images + clinical data) to enhance generalization, tackling inconsistent image quality and disease progression variability.

Crucially, we engaged clinicians to address real-world adoption barriers. Their feedback shaped our design, ensuring practicality and alignment with workflows—a key step for AI integration into routine care.

By automating early-stage AMD classification, we aim to reduce costly delays (e.g., treatments like Aflibercept at ~$2,000/dose) and preserve vision through timely intervention.

Huge thanks to teammates Mark, Rohit, and Ann, and Moorfields Eye Hospital for collaboration. Watch our prototype demo here (Due to health reasons I didn't attend the demo)!

MLOps London Meetup Insights: LLMs, AGI, and Serving Strategies

The latest MLOps London Meetup explored cutting-edge AI advancements. Dr. Jodie Burchell opened with "The IQ of AI", linking Claude Shannon's logic gates to GPT-4's AGI potential. She highlighted Francois Chollet's ARC benchmark for measuring skill acquisition and urged caution around biases in AGI systems.

Chris Samiullah unpacked open-source LLM development, tracing progress from GPT-3.5 to locally deployable models via quantization and GPU optimization. He championed RAG for accuracy and introduced DeepEval for LLM assessment, aligning with Andrej Karpathy's "limitless scaling" vision.

Ramon Perez closed with model serving strategies, contrasting Batch, Online, and Streaming approaches. He outlined deployment paradigms—embedded models (edge devices), Model-as-a-Service (collaboration-friendly), and niche Model-as-Data systems—each balancing trade-offs in scalability and maintenance.

Together, these talks underscored MLOps' rapid evolution, blending theoretical rigor (ARC, AGI ethics) with pragmatic tools (RAG, quantization) to shape AI's future.

PyData London 80th Meetup: RAG, Dependencies, and OS Debates

Victor Naroditskiy's talk on AI-driven enterprise search highlighted replacing outdated keyword methods with semantic querying. By encoding text into embeddings, Retrieval Augmented Generation (RAG) bridges queries to siloed data (Jira, Slack, etc.) while linking answers to sources—unlike ChatGPT. Key insights: prioritize fine-tuning embeddings over full LLMs for cost efficiency, and blend semantic search with filters (dates/keywords) for hybrid robustness.

A lively Python dependency panel debated tools like pip, conda, poetry, and rye. Experts weighed reproducibility, isolation, and ecosystem fit—a reminder that no "perfect" solution exists, but context rules.

Lightning Talk: Casper Da Costa-Luis shared his switch to Windows for ML workflows via WSL2, praising seamless Linux tool integration (Python stacks, CLI) within Windows' UI. Personal note: While WSL2 bridges gaps, I'll stick with native Linux for its flexibility and terminal-first ethos!

From semantic search to dependency chaos, I enjoyed the first talk most though RAG seems to be the new shiny thing!

Series of talks on career development, recruitement and growth within the tech industry. I met a lot of recruiters and made some new friends.

Adam MacVeigth from News Uk gave a talk about how the develop an Audio pipeline for text transcription using STT (Speech to Text) and TTS (Text to Speech) models using whisper on their Google Cloud infrastructure. One of the major challenges they faced was to splitting speakers and it was interesting to me because I had recently develop a similar POC application earlier for my podcast to transcribe audio to Text and make a summary, so I could relate. I won a Google Hat during the raffle draw and I presently preparing for the Google Data Engineer certification.

At the PyData London Meetup, two talks stood out for their practical insights into deploying data solutions effectively.

1. From Jupyter to Web Apps with Taipy
Marine Gosselin and Florian Jacta demoed Taipy, an open-source platform transforming Python data workflows into full-stack web apps. Designed to bypass JavaScript hurdles, Taipy integrates frontend/backend infrastructure, Plotly visualizations, and collaboration tools—directly from Jupyter or IDEs. While its rapid prototyping impressed me, I noted the absence of experiment tracking, a critical gap for iterative data science. Still, Taipy's vision to turn models into interactive apps (not just static reports) could democratize data product delivery.

2. Michael Natusch's 10 Rules to Not Fail at ML
Michael's talk resonated deeply: many teams know these rules but overlook them. Highlights:

• Avoid "ML sprinkles"—integrate cross-functional teams early.
• Prioritize MLOps before scaling, not as an afterthought.
• Choose practicality over novelty (e.g., simpler models for clearer ROI).
• Design systems for ML, not retrofitted software.

His emphasis on ethics, governance, and user-centric impact was a reminder that ML success hinges on process, not just algorithms.

Takeaways
Taipy excites but needs maturity (hello, experiment tracking!). Michael's rules? Obvious yet underappreciated—a blueprint for ML done right.

From Forecasts to Chatbots: Event Takeaways

Leonidas Tsaprounis dissected forecast evaluation, contrasting point metrics (RMSE, MAE) with distributional methods (log score, CRPS). While point forecasts target mean/median accuracy, distributional models offer probabilistic ranges—key for inventory planning. A critical insight: near the median, CRPS and log scores conflict, underscoring trade-offs in uncertainty modeling.

John Sandall's private ChatGPT demo—built with Streamlit, LangChain, and Vicuna-13B—showcased accessible AI tooling. Using llama.cpp, he optimized the LLaMA-finetuned model to run smoothly on a Macbook, proving lightweight LLM deployment is achievable.

In lightning talks:

• Casper Da Costa-Luis challenged open-source legality, highlighting accountability gaps in critical infrastructure.
• Jennifer Ding promoted London Data Week, advocating "data in public for public" via talks and projects like Citimap.

Personal reflection: While tools like Streamlit democratize AI apps, Casper's talk reminded me that open-source's legal ambiguities demand proactive governance—a nuance often overlooked in tech's rush to innovate.

How do you balance practicality with ethical rigor in your projects?

My cofounder and I were pitching our startup SKincare Pal to some UCL students to join us for an internship ion different technical roles as we we're seeking to start the development and needed some engineers on board. I was intrigued by some conversations with some students in the Masters program in AI and they we're eager to join our team. On the image I was explaining the technical details to some of the software engineering majors and the challenges we are solving at Skincare Pal with novel approach to skin care product recommendation using images.

I volunteered at PyData's London Flagship event, which is a 3 days conference on best practices in developing and shipping Data and Machine learning applications. It was rich day with speakers from startups and big tech corporations either selling a new service, demo, or trying to recruit. My main role was to register participants, chair different workshop sessions by presenting speakers and managing Q&A sessions after each talk. So much to learn and equally connected with some wonderful people in the ML community in London and from different places in the world.

This was a one-day event with different workshops and talks all going on in parallel. I chose to attend a workshop on recommendation systems using Multi-Armed Bandits because of my interest in recommendation systems, which turned out to be fascinating as I came away with an understanding of how Multi-Armed Bandits work using exploration and exploitation, and how to optimize the reward function. I also enjoyed the panel discussion with PwC on Responsible AI, though the panel did not have a clear answer to my question about making open source responsible without stifling innovation and regulating open source, which, in my humble opinion, contributes to innovation.

Sefik Ilkin Serengil (Vorboss) unpacked billion-scale facial recognition using FaceNet's deep CNNs to generate face embeddings. By measuring Euclidean distances between vectors, the model distinguishes identities. For speed at scale, Approximate Nearest Neighbors (ANNs)—leveraged by Spotify (ANNOY) and Meta (FAISS)—trade slight accuracy for efficiency, ideal for real-time applications.

Pavel Katunin and Anton Nikolaev showcased ML-driven stem cell research, automating neuron conversion via computer vision. Their open-source robotic system tracks cell differentiation, using focus measures refined by ML-trained sharpness models. A robust backend (S3 storage, metadata management, node orchestration) ensures reproducibility and scalability—key for high-throughput experiments.

In a lightning talk, Besart Shyti (Meta) shared his transition from software to ML engineering, advocating for hands-on projects and pair programming over passive learning. His mantra: "Build from scratch, iterate fast, and embrace collaboration."

Takeaway: From facial recognition to bio-optimization, PyData highlighted tools balancing scale and precision—while Besart's journey reminded us that growth often lies beyond traditional paths.

This workshop, organized by Snowflake with sponsorship from EDF, lasted one day and provided a remarkable opportunity to learn about the development of data pipelines and machine learning (ML) applications using Snowflake's platform. Over the course of three hours, I was able to create a small pipeline that ingested financial data into Snowflake's data warehouse, utilizing Snowpark—a framework that enables data processing and pipeline execution in languages such as Python, Java, and Scala. Additionally, we developed a regression model and built a client-facing application using Streamlit, which facilitated interaction with our ML outputs. During the workshop, I also had the pleasure of meeting an acquaintance in person—visible in the first picture—who had previously assisted me in troubleshooting a Docker issue I encountered while reaching out to him on LinkedIn

This meetup was quite interesting as we learned about from Alex Glaser of UK Health Security Agency how Data Science was used during the COVID lockdown period for infectious disease modeling to help in the development of response plans using Time Series Modeling and detecting symptons. The second talk was from Peter Vidos who presented ipyvizzu an open source framework for building animated data stories in Jupyter to present & sharing findings.