Vikram Lingam

I build small AI models and weird little experiments. Sometimes they work.

[Human Being @ensemblehive]

The AI stack is becoming surprisingly simple:

Run models locally → Ollama
Try open models → Hugging Face
Compare models → Artificial Analysis
Access multiple models → OpenRouter
Vector search → FAISS
RAG → LlamaIndex
Agents → LangGraph
Observability → Langfuse
Evaluation → Ragas
Build APIs → FastAPI

You don’t need 50 AI tools.

You need a small stack you actually understand.

[Human Being @ensemblehive]

A simple guide to choosing an LLM:

Need maximum reasoning → Frontier model
Need low cost → Small model
Need privacy → Local model
Need offline use → Open-weight local model
Need fast responses → Smaller / distilled model
Need your own knowledge → RAG
Need a specific behavior → Fine-tuning
Need actions → Model + tools
Need reliability → Model + deterministic code

Dont ask: “What’s the best model?”

Ask: “What’s the smallest model that solves my problem well?”

[Human Being @ensemblehive]

10 AI websites that are ridiculously useful:

https://t.co/Jba2TPvmXf → compare AI models by intelligence, speed & price
https://t.co/CexCd5YCuX → access and compare hundreds of LLMs in one place
https://t.co/iEZmkz52Jw → try thousands of AI demos, often for free
https://t.co/iDKiWJufGe → compare models blindly and see which actually performs better
https://t.co/YbFuhizh1h → run open-source AI models through simple APIs
https://t.co/b9NbZGzZ5h → research papers + implementations
https://t.co/hbevPXWYoG → discover new AI research without hunting through arXiv
https://t.co/D5xn9chZGY → find open-source alternatives to popular AI/software tools
https://t.co/BEiOYS378t → install and run many AI apps locally
https://t.co/IrgYAEWxUU → discover models you can run on your own computer

Most people only use the big AI chatbots.

The interesting stuff is often hiding elsewhere.

Bookmark this.

[Human Being @ensemblehive]

15 websites worth bookmarking:

https://t.co/qY34Eq6yMM → remove image backgrounds
https://t.co/sMSATwwymU → Photoshop in your browser
https://t.co/kP9rGfswpr → PDF, image & file tools
https://t.co/Z3vJ35vrwZ → quick diagrams & sketches
https://t.co/RXLWEubxUp → compress images
https://t.co/YDW1aS9XV3 → find alternatives to any app
https://t.co/BUTLM6EXm4 → check if your email was in a data breach
https://t.co/fUGslhQ90l → analyze suspicious files & URLs
https://t.co/OFJvHQJB89 → find where movies & shows are streaming
https://t.co/dGDQ9IdFld → install multiple Windows apps at once
https://t.co/a6szDUKxEO → remove unwanted objects from images
https://t.co/fm45GZkPCA → dozens of free PDF tools
https://t.co/hHI2TsJ6WW → build & debug regular expressions
https://t.co/D0FWXRVpT9 → create beautiful code screenshots
https://t.co/IaxB5pSdlj → the Internet Archive

The internet is still full of ridiculously useful little tools.

Bookmark this. You will eventually need one of them.

[Human Being @ensemblehive]

Mathematics is strange and beautiful at the same time

[Human Being @ensemblehive]

Free tools I would install first on a new AI dev machine:

VS Code → coding
Git → version control
Docker → containers
Ollama → run LLMs locally
LM Studio → test local models visually
Hugging Face → models & datasets
Jupyter → experiments
Postman → test APIs
DBeaver → databases
GitHub Desktop → easier Git workflow

You can build a surprising amount without paying for a single tool…✌🏻

[Human Being @ensemblehive]

Want to understand AI instead of just using it?

Start here:

Neural networks → 3Blue1Brown
Deep learning → https://t.co/TdWy7trdBp
Transformers → The Illustrated Transformer
LLMs from scratch → Andrej Karpathy
Hugging Face models → HF Course
Machine learning → Google ML Crash Course
Practical ML → Kaggle Learn
AI research papers → Papers with Code

Most of these are free.

You don’t need another $500 AI course.

[Human Being @ensemblehive]

This is pure mathematics!

Not generated by any AI!

[Human Being @ensemblehive]

Learn AI by actually building things:

Neural networks from scratch → micrograd
Transformers from scratch → nanoGPT
LLMs from scratch → build-nanogpt
Embeddings → Sentence Transformers examples
Local LLMs → Ollama
AI agents → smolagents
Model fine-tuning → Unsloth

Reading teaches you the concepts.

Breaking these teaches you how they actually work… ✌🏻

[Human Being @ensemblehive]

A simple guide to making LLM apps more reliable:

Wrong facts → RAG
Bad instructions → Better prompting
Inconsistent JSON → Structured output
Repetitive mistakes → Few-shot examples
Needs calculations → Use code/tools
Needs live information → Search/API
Needs private data → Retrieval
Too expensive → Smaller model
Too slow → Cache + smaller model
Complex task → Break it into steps

The answer to every AI problem isn’t a bigger model…✌🏻

[Human Being @ensemblehive]

you are backed by math!

[Vikram Lingam @vikram_lingam]

@aka_kosasih This is just beautiful.

[Human Being @ensemblehive]

A simple ML model cheat sheet:

Predict a number → Linear Regression

Predict a category → Logistic Regression

Complex tabular data → XGBoost

Need an interpretable baseline → Decision Tree

High-dimensional small dataset → SVM

Group similar things → K-Means

Reduce dimensions → PCA

Text similarity → Embeddings + Cosine Similarity

Images → CNN / Vision Transformer

Sequential text → Transformer

Do not start with the fanciest model or fanciest name. This is where many beginners gets trapped.

Start with the simplest one that can solve the problem. ✌🏻

[Human Being @ensemblehive]

A surprisingly useful rule when building with AI:

If you can solve it with code, use code.

If you can solve it with search, use search.

If you can solve it with rules, use rules.

Use an LLM only for the part that actually needs an LLM.

Your AI system usually gets cheaper, faster and more reliable.

[Human Being @ensemblehive]

Math becomes much more interesting when you stop treating equations as things to solve.

[
\begin{aligned}
x_i(t)&=(r_i+58\sin(0.1r_i-3t))\cos\left(0.2375i+\frac{160}{r_i+7}-1.5t\right)\
y_i(t)&=(r_i+58\sin(0.1r_i-3t))\sin\left(0.2375i+\frac{160}{r_i+7}-1.5t\right)
\end{aligned}
]
with (r_i=2.5\sqrt{i}).

This equation doesn't describe the animation.

It is the animation...

[Vikram Lingam @vikram_lingam]

Every time I learn how one part of AI actually works, it becomes less magical.

And somehow more interesting.

[Vikram Lingam @vikram_lingam]

@sundarpichai @collision @OpenRouter Will you acquire Stripe? 🧐

[Vikram Lingam @vikram_lingam]

tech people, let’s connect 👋

if you’re into:

💻 backend
🎨 frontend
⚡ full stack
🛠️ devops
📊 data
✨ UI/UX
🚀 startups
💰 SaaS
🤖 AI
🔥 vibe coding

tell me what you’re building 👇

always looking to meet more builders 🤝

[Vikram Lingam @vikram_lingam]

The more I learn about AI, the more I realize how much I still don’t know.

And weirdly, that’s the part I enjoy most.

Every new thing I understand opens up five more things I want to learn.

[Vikram Lingam @vikram_lingam]

@AravSrinivas Recently had a thought that Perplexity might be slipping out of the race. You guys are so back ✌🏻 🚀