685: Tools for Building Real-Time Machine Learning Applications, with Richmond Alake

Episode Summary

Executive Summary: Richmond Alake discussed his role as a machine learning architect at Slalom Build, emphasizing systems thinking across data engineering, MLOps, UI, and ML. He shared how his two startups—Mini PT for real-time fitness form correction and Open Speech for audio-to-content generation and analytics—use modern AI stacks. He also explained why writing and teaching accelerate learning, visibility, and career growth in fast-moving AI.

Main Topics: Machine learning architecture as systems design: Richmond defined the architect role as a high-level, cross-functional position spanning data engineering, MLOps, front end, back end, and ML integration, rather than only model building. Real-time data infrastructure at Slalom Build: He described working with Databricks and Kinesis to build scalable, streaming-ready data pipelines and emphasized that ML initiatives require mature data foundations before modeling begins. Mini PT: real-time AI fitness coaching: Richmond outlined his iOS startup using computer vision and pose estimation to track exercise form, provide audio feedback, and generate post-workout analytics for more accessible personal training. Open Speech: generative AI for content and analytics: He explained an audio-based app that generates posts, newsletters, and summaries from recordings while also analyzing speaking balance, sentiment, and filler words for podcasting and therapy use cases. Feature stores and MLOps: Richmond highlighted feature stores, especially Feast, as critical infrastructure for consistent feature definitions, real-time inference, and batch training across teams and environments. Writing and teaching as career accelerators: He argued that public writing helps retain knowledge, build credibility, and create opportunities, citing his articles, courses, and contracts with NVIDIA, Built In, Neptune AI, and educational institutions. AI pace, uncertainty, and responsible optimism: The conversation closed on how fast AI is evolving, the importance of staying current without pretending to know everything, and the need to balance progress with safety and responsible use.

Key Arguments: Machine learning architects must think in systems, not just models, because production AI depends on data pipelines, infrastructure, APIs, and user experience as much as on algorithms. Companies often attempt ML before their data and architecture are ready; real-world consulting frequently starts with data structuring, streaming, and platform maturity. Databricks is valuable because it supports large-scale, real-time data processing, orchestration across environments, SQL/Python workflows, and ML tooling like feature stores and data catalogs. Kinesis and similar streaming platforms are important for efficient real-time data delivery, even though Richmond is still learning the stack firsthand. Mini PT applies real-time computer vision to make personal training more affordable, inclusive, and accessible by giving live form correction and post-session feedback. Open Speech began as a personal productivity tool for converting podcast audio into multiple content formats, but evolved into a broader generative AI and analytics product. Open Speech’s potential in mental health requires stronger privacy, anonymization, and hosting controls than a standard content-generation app. Feature stores solve a practical MLOps problem: maintaining consistent feature definitions across teams and environments while supporting online inference and offline training. Writing is a deliberate learning tool: it forces deeper understanding, creates professional visibility, and can lead to opportunities such as books, courses, contracts, and partnerships. The AI field is moving so quickly that even experienced practitioners feel behind; continuous learning and humility are necessary to stay relevant.

Data Points: Episode number: 685 - This Super Data Science Podcast episode featuring Richmond Alake Startup count: 2 - Richmond is co-founder of Mini PT and Open Speech Personal training cost in London: upwards of £60/hour - Example used to motivate Mini PT’s affordability mission Personal training cost in New York: about $120/hour - Comparison used to illustrate how expensive live personal training can be Gym training frequency: at least 1 hour 30 minutes every day - Richmond described his routine and motivation for building Mini PT Professional experience in machine learning: about 3 years professionally - Richmond explained his rapid career acceleration Master’s duration in AI: 1 year - Part of Richmond’s educational path in the field Feature store course length: about 3 hours - His upcoming O’Reilly live training on Feast and feature stores Mini PT app platform: iOS / Swift - Core development stack for the fitness app Open Speech proof-of-concept stack: OpenAI API, Next.js, React, Firebase, Python Flask, AWS - Prototype stack for the audio-to-content and analytics product Noted salary level for cutting-edge LLM work: potentially eight-figure salaries - Mentioned in discussion of LLM ops and infrastructure talent demand Training cost reduction on AWS chips: up to 50% - Sponsor mention for AWS Trainium training-cost savings Inference cost reduction on AWS chips: up to 40% - Sponsor mention for AWS Inferentia inference-cost savings

Pivotal Quotes: "As a machine learning architect, you've got to start thinking on a system level." — Richmond Alake: He defined the role difference between architecture and traditional ML engineering "Writing is a hack." — Richmond Alake: He explained why public writing helps him learn, retain knowledge, and build visibility "Feature stores essentially allow you to take some of the headache away from redefining features or sharing features across a team." — Richmond Alake: He summarized why Feast and feature stores matter in production ML

Implications: For practitioners, success in AI increasingly depends on systems, data maturity, and MLOps—not just model skill. For careers, writing and teaching can compound expertise into opportunity. For industry, real-time, privacy-aware AI products will need strong infrastructure and responsible design.