❤️ Husband, Father 👾 AI Engineering 🧠 ML Engineering
Joined May 2018
- Tweets4.3K
- Following1K
- Followers905
- Likes9.5K
Just tried Gemini chat for the first time, trying to brainstorm through an idea, pretty painful. Took me back like four years 🤮
#ai #SuperIntelligence
Stop building one AI that does everything. Start building teams of agents that each do one thing well.
There's a quiet shift happening in how the best teams are deploying AI.
For the last two years, the instinct has been to reach for a single, general-purpose model and prompt it into submission. One assistant, infinite tasks. It's convenient — and it plateaus fast.
#AI #MultiAgentSystems #AIStrategy #MachineLearning #EnterpriseAI
Most AI agents should start as workflows.
Not because agents are bad — but because business data is messy.
If you point RAG at everything and expect great answers, you usually get inconsistent results. PDFs and Word docs are one thing, but once you get into Excel files, calculations, approvals, exceptions, and business rules, raw retrieval is not enough.
The better pattern is:
Ingest the data.
Clean it.
Structure it.
Synthesize it.
Then use RAG on top of that trusted context.
That is where agents become useful.
A workflow gives the agent the guardrails, business logic, and prepared data it needs to produce answers people can actually trust.
Don’t make the agent guess.
Make the workflow do the heavy lifting first.
#AIAgents #AgenticAI #AIWorkflows #RAG #RetrievalAugmentedGeneration #EnterpriseAI #BusinessIntelligence #DataEngineering #WorkflowAutomation #AppliedAI #AIArchitecture #LLMOps #GenerativeAI #DataStrategy #AITransformation
🤖 Made with AI
Production RAG looks less like a clever vector search and more like a data platform problem.
Freshness SLAs. Backfills. Idempotent reprocessing when you swap embedding models. Lineage from answer back to source doc.
The ML is maybe 20% of the system.
#RAG #DataEngineering #MLOps #Dagster
started Stanford CS 153 Frontier Systems last night
anjney midha's opening lecture feels more like a netflix doc than a class lol
the framing of the modern AI stack + why "context" and verifiable feedback loops decide where value accrues… gonna be chewing on that one
Day 7 / ∞ — Logistic Regression with Scikit-Learn
3 lines to fit, predict, and score a classification model.
Simple? Yes. But understanding the "why" behind the abstraction is what separates copy-pasters from practitioners.
One week down. Fundamentals are clicking.
#MachineLearning #ScikitLearn #100DaysOfML
Day 6 / ∞ — Cost Function for Logistic Regression
MSE + sigmoid = non-convex loss. Gradient descent gets stuck.
The fix → log loss:
y=1: −log(ŷ) → low confidence punished hard
y=0: −log(1−ŷ) → false confidence punished hard
Result: convex, one global minimum, gradient descent wins every time.
#MachineLearning #MLJourney #LearningInPublic
ml-visualized.com is the ML learning resource I wish had existed years ago.
Interactive visual walkthroughs of core ML concepts — gradient descent, neural nets, dimensionality reduction — all explorable, not just readable.
Perfect to have open alongside your actual work. Theory sticks when you can see it.
🔗 ml-visualized.com
Day 5 / n — Logistic Regression
The sigmoid function is the bridge from regression to classification.
Take a linear output, squash it between 0 and 1 — now you have a probability. Set a threshold (0.5), and you have a decision boundary.
The boundary is still linear. What changes is the confidence. Far from the line = high certainty. Near it = uncertain.
Same linear algebra, completely different interpretation.
#MachineLearning #MLJourney #LearningInPublic #LogisticRegression
Day 4 — SGD vs MLP on Titanic: Biggest lessons from 891 rows:
→ One feature (title extraction from names) beat every hyperparameter change
→ "Balanced" class weights dropped accuracy 6pts on a 59/41 split — not every dataset needs rebalancing
→ SGD ceiling: 79%. MLP with same features: 81.5%
→ Deeper networks lost to shallower ones. Small data can't support big architectures. Failed experiments taught more than successful ones.
#MachineLearning #MLJourney #LearningInPublic #ScikitLearn #DataScience
📊 What are some ways that companies are hosting models in production? Are most companies doing one of the following?
- Customer Integrations API "FastAPI"
- NVIDIA Triton Inference Server
- Other
Day 3 — Feature Engineering, Polynomial Regression & Learning Rate
Feature engineering = giving the model better inputs, not more parameters.
Polynomial regression = adding complexity where curves exist. Overfitting = adding it where they don't.
Learning rate too high → diverge. Too low → crawl. The right rate isn't fixed — it's scheduled.
All three boil down to one thing: right information, right scale.
#MachineLearning #MLJourney #LearningInPublic
Day 2: Regression with Multiple Input Variables
Moving beyond simple linear regression — working with multiple features and the mechanics of how models actually learn.
The focus: gradient descent, and what changes when you implement it with vs without vectorization.
Without vectorization, you loop through each feature and example, computing partial derivatives one at a time:
wⱼ := wⱼ - α · (1/m) Σ (f(x⁽ⁱ⁾) - y⁽ⁱ⁾) · xⱼ⁽ⁱ⁾
With vectorization, the same update becomes a single matrix operation:
w := w - (α/m) · Xᵀ(Xw + b - y)
Same math. Dramatically different performance. The loops make the calculus visible. The vectors make it practical.
Building the intuition before reaching for the abstractions.
📅 Day 2 / n
🔖 #MachineLearning #GradientDescent #LinearRegression #MLJourney #LearningInPublic
Day 1 of My Machine Learning Journey 👾
I'm starting my ML learning journey and sharing what I pick up along the way. Day 1 — the fundamentals.
Supervised Learning: The model learns from labeled data — inputs paired with known outputs. You give it the answer key and let it find the patterns. Classification (spam or not spam) and regression (predicting prices) are the core use cases. Label quality determines model quality.
Unsupervised Learning: No labels. No answer key. The model finds structure on its own. Clustering groups similar data points together, dimensionality reduction compresses features while keeping what matters. You don't tell it what to look for — it tells you what's there.
Gradient Descent: This is how the model actually learns. Start with random parameters, measure how wrong you are (loss function), and adjust weights in the direction that reduces error. The learning rate controls step size — too big, you overshoot, too small, you wait forever. Repeat until convergence.
Key takeaway: Supervised learning needs labels, unsupervised learning finds hidden patterns, and gradient descent is the optimization engine that makes it all work.
Day 2 coming soon.
#MachineLearning #ML #AI #LearningInPublic #SoftwareEngineering #Day1
