Joseph Eton

[Alexey Grigorev @Al_Grigor]

Once an LLM application is deployed, the work is not finished.

Now you need to understand what happens when real users start using it:

- What questions do they ask?
- Which documents are retrieved?
- What context is passed to the model?
- What answer is generated?
- Was the answer useful?
- Where does the pipeline fail?

In the next LLM Zoomcamp workshop, we'll focus on monitoring.

📅 Wednesday, June 3, 2026
🕛 12:00-13:30 CET
📍 YouTube

We'll go through a hands-on example and look at how to:

- Collect traces from an LLM pipeline
- Instrument LLM calls and retrieval steps
- Store chat history and user interactions
- Collect user feedback on generated answers
- Track answer quality over time
- Use dashboards to inspect system behavior
- Run evaluations on production traces

Before deployment, we evaluate on test data. After deployment, we use traces, feedback, and production metrics to understand what actually happens in the system. This helps us detect issues in retrieval, generation, latency, cost, and answer quality.

Register here: https://t.co/fXhQliFWgM

[Joseph Eton @EtonJoe]

@ruffydfire Hahahahahahaha one of the most difficult question.
But not too late to ask. You are in order.

[Joseph Eton @EtonJoe]

@HMRCcustomers Hi

[Joseph Eton @EtonJoe]

@HMRCcustomers hi

[Joseph Eton @EtonJoe]

@ritapurity Congratulations Prof.

[Joseph Eton @EtonJoe]

@Stephen20858296 Expect more poverty and hardship bah!

[Maziyar PANAHI @MaziyarPanahi]

🚨 Over 1 billion rows of psychiatric genetics data. Now on Hugging Face.

ADHD. Depression. Schizophrenia. Bipolar. PTSD. OCD. Autism. Anxiety. Tourette. Eating disorders.

12 disorder groups. 52 publications. Every GWAS summary statistic from the Psychiatric Genomics Consortium.

Before: wget, gunzip, 20 minutes debugging separators, repeat 50 times.
Now: one line of Python.

[Joseph Eton @EtonJoe]

@feedinkoo Im interested.

[Talk Church @churchtalkative]

CHEATCODE FOR LIFE:

1. Psalm 42 - when worry strikes

2. Psalm 112 - if you feel weak

3. Psalm 27 - when fear arises

4. Psalm 118 - if you lack confidence

5. Psalm 61 - if you need an answer

6. Psalm 41 - if you're battling sickness

7. Psalm 91 - when you need protection

[Joseph Eton @EtonJoe]

@nathmichael6 @ruffydfire Not INEC but wike...... where is your sense baba?

[#Renewed Shege ProMAX #EndSARS @prettycachy]

@OurFavOnlineDoc Don't say anything just retweet so the world can see how foolish and two mouthed Reno OMOCKERY is.

[Joseph Eton @EtonJoe]

@Oludeewon Lol 😆

[Joseph Eton @EtonJoe]

@HMRCcustomers I have an issue

[Joseph Eton @EtonJoe]

@thecableng Shaking this table...hmmmm

[Joseph Eton @EtonJoe]

@Med1_Ai AI

[Joseph Eton @EtonJoe]

@ruffydfire Lol

[Maryam Miradi, PhD @MaryamMiradi]

📄 𝗙𝗿𝗼𝗺 𝗣𝗮𝗽𝗲𝗿 𝘁𝗼 𝗣𝗿𝗮𝗰𝘁𝗶𝗰𝗲:
A new 𝗠𝗲𝗱𝗶𝗰𝗮𝗹 𝗔𝗜 𝗣𝗮𝗽𝗲𝗿 shows end-to-end 𝗔𝗴𝗲𝗻𝘁𝘀 from data to reasoning and I’ll show you how you 𝗯𝘂𝗶𝗹𝗱 𝗶𝘁 𝘆𝗼𝘂𝗿𝘀𝗲𝗹𝗳. 👇

𝗪𝗵𝗮𝘁 𝘆𝗼𝘂 𝗯𝘂𝗶𝗹𝗱:
In hospitals, data comes as notes, scans, and spreadsheets — fragmented and messy.

You’ll build an AI Agent system that turns raw hospital data into clean, private, interpretable predictions.
Powered by cooperating agents — no manual work.

Fast. Transparent. Scalable.

𝗦𝘁𝗲𝗽𝘀:

【1】 Ingestion Identifier Agent
Checks incoming data: spreadsheet, clinical note, or scan image.
Auto-detects type so the right workflow starts.

→ Think: file scanning + MIME detection + structured-data check.
→ Framework: implement inside an Ingestion Agent with CrewAI or LangGraph.

【2】 Data Anonymizer Agent
Privacy first: removes names, IDs, record numbers.
Delivers fully anonymized data ready for analysis.

→ Think: regex filters + hashing + redaction layer.
→ Framework: a local PrivacyAgent for text & image anonymization.

【3】Feature Extraction Agent
Understands what the data means.
Tables → headers like Age, Gender, Diagnosis.
Images → scan type & condition (e.g., colonoscopy → polyp).

→ Think: header parsing + embedding similarity + lightweight vision labeling.
→ Framework: MetadataAgent with pandas (tables) or simple vision modules (scans).

【4】 Model–Data Feature Matcher Agent
Chooses the best model automatically.
Text → LLM for reading/reasoning.
Tabular → structured model.
Imaging → VLM for medical scans.

→ Think: feature matching + routing logic + model registry.
→ Framework: ModelSelectorAgent that switches between LLMs, VLMs, or structured models dynamically.

【5】 Preprocessing Recommender & Implementor Agents
Recommend and apply cleaning/normalization before inference.
Tables scaled, missing values filled; images resized.
Everything optimized for the selected model.

→ Think: pandas.fillna(), StandardScaler(), torchvision.transforms.
→ Framework: PreprocessorAgent linking cleaning steps to model metadata.

【6】 Model Inference Agent (with Explainability)
Runs the model and explains decisions.
Predicts (e.g., detect a polyp, estimate risk) and shows why (features or image regions).

→ Think: SHAP values + attention heatmaps + local explanations.
→ Framework: ExplainabilityAgent connected to model outputs.

Paper: https://t.co/KLiuT9GihG

🏆 In my AI Agents Mastery training,
I teach you to build systems like this—step by step.

From Healthcare to Finance to Education,
you’ll connect these building blocks into real, working, explainable AI agents.

≣≣≣≣≣≣≣≣≣≣≣≣≣≣≣≣≣≣≣≣≣≣≣≣≣≣

⫸ꆛ Want to build Real-World AI Agents?

Join My 𝗛𝗮𝗻𝗱𝘀-𝗼𝗻 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁 𝟱-𝗶𝗻-𝟭 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 !

➠ Build Agents for Healthcare, Finance, Smart Cities & More
➠ Master 5 Modules: 𝗠𝗖𝗣 · LangGraph · PydanticAI · CrewAI · OpenAI Swarm
➠ Includes 9 Full Projects
👉 𝗘𝗻𝗿𝗼𝗹𝗹 𝗡𝗢𝗪 (𝟱𝟲% 𝗢𝗙𝗙):
https://t.co/5i2v1fIrhJ

[Joseph Eton @EtonJoe]

I just published Empowering Nottingham’s Educators: Building a Smarter Safeguarding Assistant with GenAI https://t.co/Dsj81HxuVY

[oseni rufai @ruffydfire]

If people are not held accountable for the dam issue in Maiduguri then kiss accountability good bye in Nigeria

[Joseph Eton @EtonJoe]

@HMRCcustomers Okay, I will send to them as it was an old P45

[Joseph Eton @EtonJoe]

@HMRCcustomers Hi, I want to update my tax code.
Where can I email my p45 to?