Marquardt retweeted
Do yourself a favor!
– Learn how retrieval actually works not just "chat with PDF"
– Pick one messy knowledge base people hate searching
– Clean and dedupe the docs before you embed anything
– Chunk by structure (headings, sections) not fixed character counts
– Add metadata to every chunk: source, date, owner, permissions
– Combine keyword search with vector search
– Add a reranker before anything reaches the model
– Rewrite vague user questions before you retrieve
– Retrieve less, but better. 5 great chunks beat 50 noisy ones
– Cite the exact source for every claim in the answer
– Say "I don't know" when retrieval comes back weak
– Enforce document permissions at retrieval time, not in the prompt
– Re-index only what changed, not the whole corpus
– Expire stale docs so old policies don't sound current
– Build a test set of real questions with known right answers
– Measure retrieval and generation separately
– Track how often answers match their cited source
– Log every query that returned nothing useful
– Turn every wrong answer into a new test case
– Cache embeddings and frequent answers
– Use a small model for routing, a strong one for the final answer
– Keep answer time under 3 seconds or people stop using it
– Handle tables, images and scanned PDFs, not just clean text
– Version your embeddings, chunking and prompts together
– Re-run your evals before switching embedding models
– Build a simple UI with sources people can click
– Deploy it with a live link anyone can try
– Write a README that explains the pipeline in 5 minutes
– Record a 60-second demo of it answering a hard question
– Share the code on GitHub and add it to your resume
You won't believe how many job offers you'll get.
Bookmark this list. Build it all.
Marquardt retweeted
Artificial Intelligence Tools & Their Use Cases. (save it)
TensorFlow ➜ Building scalable deep learning models for computer vision and NLP
PyTorch ➜ Dynamic neural networks for research and rapid AI prototyping
LangChain ➜ Creating AI agents with memory, tools, and chaining for complex workflows
Hugging Face Transformers ➜ Pre-trained models for text generation, translation, and sentiment
OpenAI GPT Models ➜ Conversational AI, content creation, and code assistance
Scikit-learn ➜ Classical ML algorithms for classification, regression, and clustering
Keras ➜ High-level neural network APIs for quick model development
CrewAI ➜ Multi-agent systems for collaborative AI task orchestration
AutoGen ➜ Conversational agents for automated programming and problem-solving
Jupyter Notebook ➜ Interactive AI experimentation, visualization, and sharing
MLflow ➜ Experiment tracking, model packaging, and deployment pipelines
Docker ➜ Containerizing AI apps for reproducible environments
AWS SageMaker ➜ End-to-end ML workflows with cloud training and inference
Google Cloud AI ➜ Vision, speech, and natural language APIs for app integration
Rasa ➜ Building customisable chatbots and virtual assistants
Add more in the comment 👇
Marquardt retweeted
1/ 🧵 New preprint! Do protein foundation models really understand local chemistry?
We probed them with NMR chemical shifts, per-atom reporters of chemical environment.
Short answer: they struggle exactly where the biology gets interesting 👇
[link] biorxiv.org/content/10.64898…
Marquardt retweeted
The Jacobian, the Hessian and the Wronskian are three determinants that answer three different questions. The first measures how a map changes volume, the second how a function curves, the third whether functions are independent.
The Jacobian determinant of a map F is det(∂fᵢ/∂xⱼ), the factor by which small areas or volumes change. The Hessian collects second derivatives ∂²f/∂xᵢ∂xⱼ; at a critical point in two variables, det H > 0 with f_xx > 0 is a minimum, det H > 0 with f_xx < 0 a maximum, and det H < 0 a saddle. The Wronskian of f₁, …, fₙ stacks the functions and their derivatives; if it is nonzero at some point, the functions are linearly independent.
For sin x and cos x the Wronskian is −1, so they form independent solutions of y″ + y = 0.
Marquardt retweeted
The Python community has lost one of its early contributors. 🕊️
Guido van Rossum, Python's creator, announced the passing of Tim Peters at the age of 74.
Peters was Python's second real user, creator of Timsort, author of the Zen of Python (PEP 20), and a member of the Python Software Foundation board for 13 years.
His contributions will live on in the Python ecosystem.
Rest in peace, Tim Peters. ❤️
Marquardt retweeted
I went through AI Engineer job posts on LinkedIn.
Almost every one asks for the same thing: proof you can make AI work in production not just in a notebook.
Here are 15 AI engineering projects that will make your resume stand out:
• RAG chatbot over 1,000+ real documents with source citations
• Agent that books, cancels and reschedules using real APIs
• AI coding assistant that writes tests for an existing repo
• Voice agent that answers calls and logs every conversation
• Model router that sends easy questions to a cheap model and hard ones to a smart one
• Semantic search engine for a real product catalog
• Document extractor that turns messy PDFs into clean JSON
• Fine-tuned small model that beats a big model on one narrow task
• LLM cost tracker showing spend per user and per feature
• Caching layer that cuts repeat API calls and latency
• Agent memory system that remembers users across sessions
• Evals suite that catches regressions before every deploy
• Tracing dashboard that shows every step an agent took
• Fallback system that retries, switches models and never shows users an error
• Red-team script that tries to jailbreak your own app
Pick 2–3.
Ship them to real users.
Post the numbers: cost, speed, accuracy.
That's your resume.
Bookmark this and send it to someone trying to break into AI engineering.
Marquardt retweeted
un tip para reducir la latencia de tus agentes:
aprovechá las llamadas en paralelo a las tools.
por ejemplo, un agente que investiga un bug necesita:
-> consultar logs.
-> buscar errores en Sentry.
-> revisar los últimos commits en GitHub.
si hace cada consulta por separado, pierde tiempo esperando respuestas.
como son independientes, puede solicitar las 3 al mismo tiempo y tu backend ejecutarlas en paralelo.
ojo, si una depende del resultado de otra, tenés que respetar ese orden.
cuando tu agente trabaja con varias APIs, esto te ahorra bastante tiempo de ejecución.
Marquardt retweeted
"Differential Equations for Engineers" is a free university-level textbook that offers a comprehensive introduction to differential equations and their applications.
The book contains over 460 pages covering first-order differential equations, higher-order linear equations, systems of differential equations, Laplace transforms, Fourier series, partial differential equations, nonlinear systems, power series methods, and much more.
It combines mathematical theory, worked examples, and hundreds of exercises, many with solutions, making it a useful resource for engineering and mathematics students, as well as anyone looking to review the fundamental concepts of differential equations.
jirilebl.github.io/diffyqs/d…
Marquardt retweeted
my phd cohort had an ongoing note we added to when we read a paper than mentioned a new "ome"
Marquardt retweeted
When cells grow together, they often align with their neighbors, forming liquid crystal-like patterns known as “nematic order.” But that order isn’t perfect, and at certain points, the alignment can break down, creating “topological defects”: small regions where cells swirl or fan out instead of lining up cleanly.
And according to research published in Science, what looks like a defect can be used strategically as built-in instructions for shaping larger structures. Instead of growing upward, cells can store and release stress in certain places along a previously flat sheetlike structure, forming a bowl, a ridge, or a fold.
Learn more: scim.ag/4j9XVjL #OrigamiDay
Marquardt retweeted
cada vez veo más búsquedas de AI Engineer.
¿qué hace alguien en ese puesto?
el alcance cambia según la empresa, pero en productos con IA suele ser quien toma modelos existentes y construye soluciones sobre ellos.
en la práctica:
1) identifica qué problema vale la pena resolver con IA
2) elige el modelo y define qué contexto necesita
3) lo conecta con datos y APIs para que pueda ejecutar tareas
4) arma evals para medir qué tan bien funciona
5) lleva todo a producción y monitorea errores, costos y tiempos de respuesta
por ejemplo, Anthropic busca un Applied AI Engineer para ayudar a empresas a integrar Claude en sus productos. incluye diseñar arquitecturas, construir prototipos, desarrollar evals y acompañar la implementación hasta el deploy.
me parece un puesto muy interesante si venís de desarrollo y querés meterte en IA. mucho de lo que ya sabés sigue siendo clave: APIs, bases de datos, testing, seguridad.
y además te obliga a entender bien el problema de negocio, decidir dónde puede aportar un modelo y comprobar si realmente lo resuelve.
Marquardt retweeted
なんと！
｢AIが設計したタンパク質は、臨床用粒子より1,000倍以上優れたmRNAの送達を実現｣
Nature
･ AIで自然界にない構造の運搬体を100種類以上作製
･ 培養細胞で比較した脂質ナノ粒子より導入効率1000倍超
･ 筋ジストロフィー患者由来細胞やブタの筋肉で遺伝子編集
nature.com/articles/s41586-0…
Marquardt retweeted
Linear algebra wasn't built in a day. It took nearly 200 years of ideas, from Leibniz's determinants in 1693 to Peano's formal definition of vector spaces in 1888.
Along the way, Cramer found a way to solve linear systems, Gauss developed elimination methods, and Sylvester gave matrices their name.
One equation captures much of this journey: Ax = b. Behind those three symbols lies a mathematical language that now powers computer graphics, engineering, AI and modern physics.
Marquardt retweeted
some of the most influential papers in AI/CS are surprisingly short.
papers you can read in one sitting:
MapReduce (2004): 13 pages, how Google processed massive datasets across thousands of machines.
The Google File System (2003): 15 pages, how distributed storage works at scale.
Adam (2014): 9 pages, the optimizer behind countless deep learning models.
LoRA (2021): 8 pages, fine-tuning large models using low-rank updates.
Dropout (2012): 2 pages, a simple trick to reduce neural network overfitting.
A Mathematical Theory of Communication (1948): foundational information theory, but substantially longer.
Marquardt retweeted
cuando trabajás con agentes, hay varias formas de repartir y organizar el trabajo.
estos 7 patrones aparecen bastante:
1) agent loop → un agente ejecuta una acción, observa el resultado y decide cómo seguir.
2) prompt chaining → el trabajo se divide en pasos: el resultado de uno es la entrada del siguiente.
3) routing → un pedido se dirige al modelo o agente más adecuado para resolverlo.
4) parallelization → varios agentes resuelven tareas independientes al mismo tiempo.
5) orchestrator-workers → un agente principal delega subtareas a otros agentes y reúne sus resultados.
6) evaluator-optimizer → un agente genera una solución y otro la revisa. el primero usa ese feedback para mejorarla.
7) handoffs → un agente le transfiere el control a otro, que queda a cargo de continuar la tarea.
lo interesante es que podés combinarlos: uno reparte el trabajo, varios lo resuelven en paralelo y otro revisa el resultado.
Marquardt retweeted
The next generation of GenAI Engineers won't be hired because they know how to wrap an LLM API. They'll be hired because they can build the infrastructure that makes AI reliable, fast and scalable. If you want to stand out in 2026, these are 15 systems worth building.
Start with the architecture of real AI applications. Build an asynchronous SSE streaming gateway, a real-time voice and vision perception loop, a generative UI with model fallback, a durable multi-agent orchestrator, and a self-correcting RAG engine. These projects prove you understand how AI moves beyond a chat window and becomes part of an actual product.
Then tackle the problems that make agents fail in production. Implement grammar-constrained JSON outputs, trajectory-based CI regression tests, an LLM trace replay debugger, and an inter-agent security layer that protects PII and blocks prompt injection. Add a cost kill-switch to stop runaway loops before they turn into expensive outages. Reliability isn't a feature you add later. It's part of the architecture.
Performance is where engineering judgment really shows. Build a KV-cache prefix-sharing proxy, a shadow-routing comparator for testing new models safely, and a WebGPU inference fallback that keeps applications running when cloud APIs fail. Measure latency, token consumption, failure rates and cost per task. A demo proves that something works once. Production engineering proves it keeps working under pressure.
Finally, turn your engineering work into public evidence. Build a DPO feedback pipeline that learns from real user preferences, then publish three serious GenAI teardowns with architecture diagrams, reproducible benchmarks and cost analysis. Don't try to complete all 15 projects just to collect GitHub stars. Pick a few, deploy them, measure what breaks and explain the trade-offs. Most people learn how to call models. The engineers who stand out know how to make entire systems work. Bookmark & Repost.
