← Back to Workshop Hub MODULE 1 · LLM FOUNDATIONS

How LLMs Read, Think & Write

A plain-language tour of how Large Language Models like the ones powering AgentSea actually work — from your healthcare query to the AI's response. No coding required.

📖 ~12 min read ⚡ Interactive 🏥 Healthcare Context L100 Foundational

🤔 What is a Large Language Model?

A Large Language Model (LLM) is a type of AI that has read enormous amounts of text — books, articles, websites, research papers — and learned the statistical patterns of how language works. The "language" in healthcare looks like procurement contracts, MOH circulars, vendor proposals, leave forms, and meeting minutes.

Think of an LLM as a very fast, very well-read junior staffer that can read, summarise, draft, classify, and reason — but always under your supervision. It doesn't understand the way humans do; it predicts the most likely next word given everything that came before.

🏥 Healthcare Analogy
An LLM is like a junior staff member who has read every memo, policy circular, and vendor proposal ever written in your organisation — but who has never actually worked a day at the hospital. Brilliant at pattern-matching language. Useless at clinical judgement. Always cite, always verify.
🧠

Pattern, Not Understanding

It learned what words usually follow other words. There's no "thinking" happening — just very sophisticated next-token prediction.

📚

Trained on Public Text

Mainly internet, books, code. It has NOT read your private hospital records, MOH internal memos, or staff handbook — unless YOU give it that context.

⚖️

Confident ≠ Correct

An LLM will sound confident even when wrong. Hallucinations happen. For healthcare, this means: always verify before acting on AI output.

🔁

Stateless by Default

Each session starts fresh. It doesn't remember last week's vendor query unless that context is re-provided in the prompt.

💡
The mental model that works: An LLM is a prediction engine for text. You give it a healthcare scenario; it returns the most probable continuation based on similar patterns it has seen. That's incredibly powerful for summarising policies and drafting emails — but never for clinical decisions, dosage calculations, or any judgement that requires real-world stakes.

🚀 Why is this a Big Deal Now?

LLMs aren't new in concept — but three things changed around 2022 that turned them into productivity tools you can actually use at work:

📈

Scale

Models grew from millions to billions of parameters. Capability emerged that smaller models simply didn't have.

💬

Conversational UI

Chat interfaces (and platforms like AgentSea) made LLMs accessible to non-engineers. No code required.

🛠️

Tool Use & Agents

Modern LLMs can now read documents, call APIs, and act — turning conversation into useful work.

🔗 The 4-Step LLM Pipeline

Every time you send a message to AgentSea, your text travels through these four steps. Click any step to learn what happens — or hit auto-play to watch the whole pipeline animate.

INPUT STEP 1 · TOKENIZE STEP 2 · EMBED STEP 3 · TRANSFORM OUTPUT 📝 Your Text "Summarise this vendor proposal" ✂️ Tokens ["Sum","mar","ise", "this","vendor"...] 📊 Embeddings [0.42, -0.13, 0.88, 0.07, …] 🧠 Transformer Attention + prediction 💬 Reply draft
📝
Step 0 · Your Text
You type a request into AgentSea — for example: "Summarise this vendor proposal and flag risks." This is just plain text. The LLM can't process letters directly.

Why is this useful to know?

You don't need to memorise the steps — but understanding that the LLM works on numerical representations of your text (not the words themselves) explains a lot:

What LLMs Do Well

An LLM is brilliant at language tasks — anything where you take text in, transform it, and get text out. Here's what works reliably for healthcare administration:

CapabilityExample in Synapxe context
📋 SummarisationCompress a 30-page vendor proposal into a 1-page executive summary with bullet-pointed risks.
🔍 ExtractionPull key terms from a contract: vendor name, contract value, SLA percentage, expiry date.
✍️ DraftingGenerate a first-draft email to a vendor declining their proposal — polite, formal, with reasons.
🔄 ReformattingConvert messy meeting notes into a structured action-items table with owners and deadlines.
🏷️ ClassificationRead incoming emails and classify by urgency: Urgent / Routine / FYI.
🌍 TranslationTranslate Mandarin/Malay/Tamil patient feedback into English summaries.
Q&A over docsAnswer "what is our policy on remote work?" by reading your staff handbook.
💡 BrainstormingGenerate 10 possible names for an AgentSea agent. Suggest interview questions for a vendor.

⚠️ What LLMs Do Not Do Well

LLMs are confidently wrong in predictable ways. Knowing the failure modes helps you decide when to trust the output and when to verify.

🎭

Hallucination

The model invents plausible-sounding facts (fake citations, made-up vendor names, wrong dates). Always cross-check against the source document.

🧮

Exact Math

LLMs are unreliable for arithmetic on long numbers. For budget calculations, use Excel or have the LLM call a calculator tool.

🩺

Clinical Judgement

An LLM cannot diagnose, prescribe, or make patient-care decisions. It is not a regulated medical device. Use only for administrative work.

📅

"What's Today's Date?"

The model has a training cut-off. It doesn't know recent MOH circulars, today's date, or anything that happened after its training data ended.

🔐

Memory of Past Sessions

Without explicit context, each AgentSea session is fresh. It won't remember Tuesday's vendor discussion on Thursday unless you tell it.

📊

Counting & Listing All

LLMs sometimes miss items in long lists or double-count. For "give me all 47 vendors", verify the output matches the source.

⚠️
The golden rule for healthcare staff: The LLM drafts, you decide. Treat AI output the way you'd treat work from a brand-new intern — useful, fast, but never published or actioned without your review.

🛡️ Why AgentSea is Safer than Public ChatGPT

Public AI assistants train on whatever you type. AgentSea is different — and that difference matters in healthcare:

ConcernPublic ChatGPTSynapxe AgentSea
Data privacyYour prompts may be used for trainingStays within Synapxe AWS account · PDPA-aligned
GuardrailsGeneric content filtersSystem prompts you can customise per agent
Audit trailLimitedSessions logged · who asked what
Document handlingFiles sent to OpenAI serversDocuments stay in Synapxe S3 (partitioned by user)
Tool useGeneric pluginsCurated skills: Doc Analysis, Doc Gen, Outlook

🏥 Real Healthcare Workflows You Can Replace Today

Not every task is a good fit for AgentSea. The sweet spot: document-heavy administrative work that's repetitive, time-sensitive, and tolerant of a "draft for review" pattern. Click any scenario to see how it maps to AgentSea skills.

🏥

Vendor Proposal Triage

30-page proposal → 1-page summary with risks, costs, SLA gaps. Saves ~2 hrs per proposal.

📋

Meeting Minutes → Actions

Raw transcript → action items with owners, deadlines, and follow-up email drafts.

💰

Budget Variance Reporting

Monthly financial → highlight variances >10%, generate executive commentary in plain English.

📧

Email Inbox Triage

Outlook integration → classify by urgency, draft routine replies, summarise threads.

📜

MOH Circular Compliance

New MOH circular → compare against your internal policy → flag gaps + draft action items.

🔄

Leave / Claims Processing

Leave form → validate against policy → generate approval or rejection with reasoned justification.

🏥
Vendor Proposal Triage
Today: Procurement officer reads 30-page proposal · 2 hrs · highlights manually. With AgentSea: Upload PDF → "Summarise key terms, list risks, compare against our SLA targets." → 1-page draft in 90 seconds. Officer reviews and edits.

🚦 When NOT to Use an LLM

Some tasks look AI-friendly but aren't. If you see these patterns in your work, route them away from AgentSea:

Task typeWhy LLM is the wrong toolBetter tool
Patient diagnosisNot a medical device. Liability + safety risk.Clinician judgement
Medication dosageHallucination risk. Calculation reliability.Validated drug-info system
Exact financial totalsNumber reasoning is unreliable.Excel · SAP · validated finance system
Regulatory submissionsHallucinations could invalidate filings.Compliance team review · LLM as draft only
Patient PII handlingDon't paste PII into prompts unless the platform is approved for it.De-identified data only
Long lists requiring 100% recallLLMs miss items in long enumerations.Database query · spreadsheet
Quick test: "If the AI is wrong here, what's the worst that happens?" If the answer is "someone gets harmed or we get fined" → don't automate it. If the answer is "I waste 2 minutes re-doing it" → great fit.

📈 Productivity Gains You Can Expect

Realistic time savings reported by similar healthcare admin teams. These assume the staff member still reviews the output — which is exactly how you should use it.

WorkflowBefore AIWith AgentSeaSaved
Vendor proposal review2 hrs30 min (90 sec draft + 28 min review)~75%
Meeting minutes → actions45 min10 min~75%
Email triage (50 emails)60 min15 min~75%
Policy gap analysis3 hrs45 min~75%
Monthly KPI commentary2 hrs40 min~67%

Across a 50-person admin cohort doing 2 of these workflows per week each, that's roughly 4,000 hours of capacity returned to the team per year — time that goes back into higher-value work.

🧭 Where this Workshop Goes Next

You now know what an LLM does and why it's useful. The next four explainers go one layer deeper into how — with the same plain-language, healthcare-context approach.

Key Takeaways

🎯 Ready for the Capstone?

The Agent Design Canvas exercise puts everything together: pick a workflow from your team, design the agent, write the system prompt, and submit for AI scoring on the leaderboard.

🤖 Start Agent Design Canvas →