Felix Sam Nanor

AI/ML Engineer | Founder and CEO of Nanor Lab | Content Creator | Robotics Enthusiast

[harpreet @DataScienceHarp]

your gaussian splat looks photorealistic from the trajectory you trained it on. move the camera off that path and the geometry falls apart

this barely gets measured because the ground truth has to be more accurate than the thing you're scoring. that means dragging a survey-grade scanner around the site for days

oxford robotics institute did it for six oxford landmarks. 24 sequences, 125,000 m², a handheld rig with three synchronized fisheye cameras, a 64-beam hesai lidar and an imu, and a leica RTC360 scan of every site as the reference — 1.9mm accurate at 10m, with the trajectories registered at 1-2cm

the novel-view test images aren't held-out frames from the training path. they're a different walk through the same site facing a different direction. that's the part that breaks splats

i packed six episodes into mcap so you can scrub all three cameras, the lidar, the imu and the slam pose on one timeline in fiftyone, with lidar depth painted onto every frame

checkout the dataset here: https://t.co/CdQALOVIWc

it's running as a live space too, nothing to install: https://t.co/IgQ8Zwt3Uc

[Felix Sam Nanor @techwithnanor]

As AI engineers, we spend all day training models. 🤖

Sometimes, you need to train the model AND train the body. 😂💪

Close the laptop.
Push some weights.
Let the GPU rest and let the CPU become the muscle. 🏋️‍♂️😂

Debug life. Deploy gains.

[harpreet @DataScienceHarp]

falling snow shows up as thousands of fake objects in lidar point clouds. radar barely notices it's snowing

most self-driving datasets are shot on sunny days in california or phoenix. none of them show what happens once the weather turns

Boreas is UTIAS's answer: 128-beam lidar, 360 degree radar, and 5MP camera driving the same Toronto route for a full year through sun, rain, and snow.

326,180 3D boxes for cars, pedestrians, and cyclists across 7,111 labeled frames

loaded as native mcap in fiftyone so you can scrub camera, lidar, and radar on one synced timeline, and watch the 3D boxes render live on the point cloud and project onto the camera and radar images

checkout the dataset here: https://t.co/KivCRtDndW

or get hands-on with this hugging face space: https://t.co/7WyTZUPj2d

[Felix Sam Nanor @techwithnanor]

Me: swears I will never get emotionally attached to lines of code

Hugging Face : releases a new open-source model with a cute name and an emoji

Me: 🥹👉👈 installs it, get more GPU cloud compute, and starts treating it like a beloved family pet

[Felix Sam Nanor @techwithnanor]

Why Most RAG Systems https://t.co/yjEtUBP4d1’s Not Always the Retrieval.
A lot of RAG applications don’t fail because the architecture is poorly built.

Sometimes, the biggest problem is the user query itself.
If the query is vague, incomplete, or poorly formulated, even a good retrieval system can return the wrong context.

Here are a few techniques I use to improve RAG retrieval:

1. Clean and rewrite the query
Before searching, compare the current query with the conversation history.
Resolve things like pronouns, missing context, and ambiguous references.
For example:
What about the second one?
That query alone is almost useless for retrieval.
But with conversation history, it can be rewritten into something like:

What are the limitations of the second RAG architecture we discussed?
Now the retriever has a much better query to work with.

2. Use Hybrid Search

Don’t rely only on vector similarity.
Combine:
Vector search captures semantic meaning.
Keyword/BM25 search captures exact terms, names, IDs, technical keywords, and phrases that semantic search may miss.
The vector search understands what the user means.
Keyword search understands what the user explicitly said.
Combining both usually gives you a stronger candidate set.

3. Use Metadata Filtering
Your chunks should carry useful metadata.
For example:
document_type: research_paper
date: 2026-07-15
topic: RAG
source: technical_blog
Then you can filter retrieval before or during search:
Find information about RAG from documents published after July 2026.

Instead of searching the entire vector database, you can restrict the search to relevant metadata.

4. Rerank the Results
Hybrid retrieval can return a large number of potentially relevant chunks.
That doesn't mean all of them belong in the final context.
A reranker takes the user query + retrieved chunks and scores how relevant each chunk is to the query.
Then you keep only the top-k results.

So the pipeline becomes:

User Query → Query Rewriting → Hybrid Retrieval → Metadata Filtering → Reranking → LLM

The important idea is that RAG is not simply:
Embed → Vector Search → LLM
Good RAG systems treat retrieval as a pipeline.

The better your query understanding, candidate retrieval, filtering, and ranking are, the better context you ultimately give the LLM.

And sometimes, improving the retrieval pipeline will give you a much bigger improvement than changing the LLM itself.
How do you currently build RAG pipelines ?

[Felix Sam Nanor @techwithnanor]

No one is literally talking about how reviewing AI generated code is harder than writing the code manually 😬😬😩😩

Nothing is free indeed 😢😌

[Felix Sam Nanor @techwithnanor]

In this scenario, the router acts as the AI agent. It analyzes the user’s query, breaks it down into a list of tasks, and determines which tools are required to complete each task. The agent then invokes the appropriate tools based on the task list.

Once a tool returns its results, the agent validates the output to ensure it aligns with the user’s original query. If the results are incomplete or irrelevant, the agent refines the tool request and retries until it obtains a satisfactory response.

Tool selection is entirely driven by the user’s intent. For example, if the query requires up-to-date or real-time information, the agent prioritizes web search. In many cases, it combines multiple tools such as web search and a database or vector store to gather comprehensive information before generating the final response. When you give it access to the right tools base on the problem you are trying to solve it will definitely do a better job. Just that it costs more than the classical RAG

[Felix Sam Nanor @techwithnanor]

Classic RAG always retrieves once and answers.

Agentic RAG lets the AI decide: → search your private docs
→ search the live web
→ or do both
I built a full tutorial with LangChain + LangGraph

Code: https://t.co/IJJbLAXB2W
Watch here: https://t.co/vAT4LtzXV4
#agenticRAG #langChain #AIAgents

[Tech Watt @techwattai]

Welcome to August!

A new month means new opportunities to learn, innovate, and build the future.

At TechWatt, we're equipping young minds with Robotics, AI, Coding & Future Tech skills.

Let's make this month count!

#TechWatt #Robotics #AI #Coding #STEM #FutureTech

[Felix Sam Nanor @techwithnanor]

Classic RAG always retrieves once and answers.

Agentic RAG lets the AI decide: → search your private docs
→ search the live web
→ or do both
I built a full tutorial with LangChain + LangGraph

Code: https://t.co/IJJbLAXB2W
Watch here: https://t.co/vAT4LtzXV4
#agenticRAG #langChain #AIAgents

[harpreet @DataScienceHarp]

hey look it’s me!

[Tech Watt @techwattai]

Give your child a head start in Robotics & AI! Build confidence, creativity, and future-ready skills through hands-on learning.

📅 Date: 1st August 2026
🕙 Time: 10:00 AM – 12:00 PM

👉Ages 8–18

🔗 Register: https://t.co/cyxnPQUvW4

#Robotics #AI #STEM #Bradford #TechWatt

[harpreet @DataScienceHarp]

whatever happened to the paligemma models? i like them. i want more.

[Felix Sam Nanor @techwithnanor]

The future of AI agents is not about adding more agents, it’s about building smarter agents.

More agents ≠ more intelligence.

Every additional AI agent comes with more complexity, more orchestration, more latency, and higher costs.

The goal is not a team of 100 agents. The goal is the right agent with the right tools, memory, and autonomy to solve the problem efficiently.

AI engineering is becoming an optimization game: capability vs cost.

Always use evaluation tools like LangSmith, Langfuse, and other to evaluate important metrics before deploying Agentic workflows.

[Felix Sam Nanor @techwithnanor]

@anycrawler Yeah kinda makes sense 💯

[Felix Sam Nanor @techwithnanor]

Most AI demos look smart until you ask something that happened yesterday.

That is the problem. Large language models are frozen in time.
They do not know what is live unless you give them a way to search.

In my new YouTube tutorial I build a real Web Search AI Agent with LangChain, GPT, and Serper from scratch. You will see how the agent decides when to search, calls a custom tool, and returns fresh answers you can actually trust.

If you are learning AI agents or building apps that need current information, this one is for you.

GitHub repo: https://t.co/HWqaXHv8FH

Full video: https://t.co/xo8LEFkGAh
#LangChain #AIAgents #MachineLearning #Python #OpenAI #BuildInPublic

[Felix Sam Nanor @techwithnanor]

@gaagana_ 😂😂😂😂

[Felix Sam Nanor @techwithnanor]

AI interviewer is becoming the new standard for most companies now. Will you allow yourself to take AI interviews? #techjobs #techinterviews #aiinterveiws #ai

[Felix Sam Nanor @techwithnanor]

AI Agent vs Agentic AI — finally explained simply

You've heard both terms. Maybe you've used them interchangeably. But they're not the same.

In this video I break down:
✅ What an AI Agent actually is
✅ What makes Agentic AI different
✅ Single agent vs multi-agent architecture
✅ When to use which approach

Perfect if you're learning AI, building with LLMs, or just trying to keep up with the hype.

🎥 Full video: https://t.co/yZX2BIYMKx

#aiagent #agenticai #artificialintelligence #llm #machinelearning #generativeai #tech #coding #aieducation #learnontiktok #techtips

[OFX @ofxproduction]

Lamine Yamal have been outstanding in the World Cup so far. #lamineyamal #worldcup #spain #spainvsfrance