Edd Mann

Edd Mann

@eddmann.com

Developer https://eddmann.com/ Podcast https://compiledconversations.com/

62 followers45 following188 posts

Pinned

i built a game boy emulator in #php over the weekend - mostly from my phone, with the help of #claude’s remote agents. 🤯 here’s how it went 👇

Ep. 19: From Linguistics to Large Language Models with Chris Brousseau i've been wanting to do an episode that really digs into what #LLMs actually are under the hood, and Chris's journey from linguistics into ML made this the perfect conversation to do it 👇 🎧 compiledconversations.com/19/

compiledconversations.comFrom Linguistics to Large Language Models with Chris Brousseau - Compiled ConversationsChris Brousseau, co-author of LLMs in Production and VP of AI at Veox AI, joins us to peek under the hood of large language models. We explore the journey from linguistics to machine learning, what…
View full thread

he's deployed billion-parameter models at mastercard and jp morgan chase, and co-authored LLMs in Production (which i'm a huge fan of). we dig into the gap between demos and production systems, and the operational challenges that make deploying at scale uniquely hard. 4/5

Ep. 19: From Linguistics to Large Language Models with Chris Brousseau i've been wanting to do an episode that really digs into what #LLMs actually are under the hood, and Chris's journey from linguistics into ML made this the perfect conversation to do it 👇 🎧 compiledconversations.com/19/

compiledconversations.comFrom Linguistics to Large Language Models with Chris Brousseau - Compiled ConversationsChris Brousseau, co-author of LLMs in Production and VP of AI at Veox AI, joins us to peek under the hood of large language models. We explore the journey from linguistics to machine learning, what…

we start with how a graduate seminar on python and machine translation in 2017 pulled chris from linguistics into NLP. studying how humans create meaning in language for years turns out to be the perfect background for understanding where LLMs shine and where they completely fall apart. 2/5

from there we go deep on the fundamentals - transformers, tokenization, embeddings, context windows, the training pipeline from self-supervised learning through RLHF to fine-tuning. 3/5

we dig into socio-technical alignment - how domain models drift from business language over time (semantic drift), and Conway's Law shaping your architecture. Architecture Modernization Enabling Teams (AMET) help bridge these gaps between business needs and technical reality. 4/5

we end with how AI tools like Claude Code are transforming architecture work - code analysis, refactoring, living documentation. Nick shares his guardrails approach: linting, test coverage, complexity limits for agents, and multi-agent workflows. 🎧 compiledconversations.com/18/ 5/5

compiledconversations.comArchitecture Modernization with Nick Tune - Compiled ConversationsNick Tune, author of Architecture Modernization and Staff Engineer at Payfit, joins us to discuss the challenges and strategies of modernising legacy systems. We explore the four pillars of…

we start by reframing legacy systems - they're not failures, they're successful products being asked to solve problems they weren't designed for. Nick introduces his four-pillar framework: business strategy, design discovery, architecture, and execution. 2/5

one key concept is "Death Valley" - the dangerous hybrid state during migration where you're running old and new systems in parallel, keeping data stores in sync. Nick also explains the "month three milestone" - balancing analysis with showing demonstrable progress. 3/5

Edd Mann reposted

I had the please to speak with @eddmann.com about event sourcing lessons learned from 10 years of production and a bunch of related patterns and concerns that have surrounded this architectural choice. I'm glad that we were able to cover a lot of non-introductory material here.

Edd Mann@eddmann.com

Ep. 17: Event Sourcing with Shawn McCool @shawnmc.cool shares how event sourcing fundamentally changed how he thinks about modeling software. if you've felt stuck in the "way we've always done it" mindset, this one's for you 👇 🎧 compiledconversations.com/17/ 1/5

a practical pattern that eliminates versioning and data retention headaches: design aggregates that exhaust quickly. at his payment processing company, event streams terminate within 30 minutes. no long-lived aggregates = no painful migrations, no GDPR nightmares. 4/5

we also cover testing with given-when-then patterns, synchronous vs asynchronous projections, and why DDD isn't a methodology - it's a pursuit of understanding your domain. lots of practical wisdom from someone who's been building these systems for over a decade. 🎧 compiledconversations.com/17/

compiledconversations.comEvent Sourcing with Shawn McCool - Compiled ConversationsShawn McCool returns to discuss event sourcing in depth. We explore what event sourcing really means, how it differs from event-driven architectures, the relationship with CQRS and DDD, and practical…

shawn's journey started around 2012 while working on laravel.io when a friend introduced him to domain events. that discovery pulled him deeper into the DDD community and eventually led to creating Event Sourcery - a free video course on domain modeling and event sourcing. 2/5

the core insight: events aren't just for messaging - they become the actual source of your model state. we dig into how this differs from event-driven architectures, the relationship with CQRS, and why capturing intent matters more than capturing CRUD operations. 3/5

spent some time exploring cli + agent skill as an alternative to the mcp servers i built a few months back. built whatsapp, strava, and garmin-connect clis to explore this space. give an llm a discoverable, well-documented cli - it can do a lot! 1/2 👇

we dig into the realities of solo bootstrapping - open core business models, building for developers, pricing traps, and why talking to customers matters more than shipping features. 4/5

we start with his time at aws - 2+ years helping teams with eventbridge and distributed systems. the same problem kept coming up: event architectures are 10-15 years behind api docs when it comes to governance and discoverability. 2/5

so he left to build event catalog full-time. what started as a christmas side project is now used by devs, architects, business analysts, and product owners - each finding different value in the same tool. 3/5

View full thread

❄️ snowball slingshot aim and launch snowballs to knock 5 ornaments off the christmas tree. angry birds vibes prob. my fav ❤️

View full thread
View full thread

he also talks about _eurekas per second_ as a design goal for learning, maximizing those aha moments per unit of time. that's a metric i can get behind 🤯

the whole feeling is less "ai will replace you" and more "what if everyone had a perfect tutor?" nice counterweight to the doom takes. worth the listen.

karpathy's take: in a post-agi world, education becomes less about climbing the economic ladder and more about... entertainment? cognitive fitness? interesting reframe.

he describes learning as going to the gym, for your brain. not easy, not "fun", but deeply satisfying. the mental equivalent of a proper workout.

smooth animations 😎...

Edd Mann@eddmann.com

my wife wanted a simple 4-8 breathing app for her apple watch. the built-in one doesn't let you customise inhale/exhale timing. so ofc i spent my lunch break building one from scratch 😅 👇 ⏱️ github.com/eddmann/Ovlo

View full thread

also built an ios companion app because why not. 😅 can start/stop sessions remotely and tweak settings from your phone. watchconnectivity is actually pretty nice to work with.

swiftui + swift concurrency made this surprisingly quick. the breathing engine is an actor so state management stays clean. whole thing is open source if anyone wants to check it out! ⏱️ github.com/eddmann/Ovlo i'm lovin' these lunch time builds.

the watch app controls everything - set your inhale duration, exhale duration, session length. smooth animated circle guides you through each breath. haptic taps on transitions so you don't even need to look at it.

with the recent opus 4.5 release, claude now tracks sonnet-only usage instead of opus. just published a new release of claudemeter to reflect that. also added a hover tooltip that shows the exact reset timestamp in your local timezone. github.com/eddmann/Clau...

View full thread

training + evaluation essentials: loss functions, gradient descent, learning rates, proper train/val/test splits, and k-fold cross-validation. plus the metrics you actually need: precision, recall, f1, auc, confusion matrices.

we kick off with ensemble learning - why combining models beats a single model. random forest, xgboost, sampling strategies, and how ensembles reduce overfitting by design.

then into neural networks: how layers of linear + nonlinear transforms learn patterns from raw data. we trace the evolution: cnns, rnns, transformers + attention, and how that leap unlocked llms.