AI-Native Solution Engineering Notes | Sprint 3

~2 min · Why this page exists

I could hand you a finished solution and teach you to analyze it. But that skips the part that matters — the process. What you see here is real: real dead ends, real moments where I didn't know what to do next, real iterations. The UMPIRE steps map to your assignments.

— Prof. Sathya

Video

Starting from Zero & Building the Map

Open Interactive Guideline Map →

Download the JSX for your own work: mdro-map.jsx

AI session: View the ChatGPT conversation →

Handwritten Reflection

What I Noticed & The Push/Pull Insight

Open Handwritten PDF →

Covers: how the touchpoint analysis reveals two failure modes — “I know but I forget” vs. “I don’t know and can’t look it up” — and why that leads to two different product hypotheses.

Framing informed by Tim Harford's Adapt — complex problems in unfamiliar domains can't be solved by planning, they have to be solved by experimenting.

Key Points / Takeaways

Slow down. I am slowing down, not trying to get to a solution fast, but trying to get the direction right by asking more first-principle questions. Handwritten reflections are helping me to slow down.
Know yourself. I am trying to tap into my own understanding of myself and my experience. I am trying to hold myself to a high standard of recognizing what I don’t know and what mistakes I might make.
Iterate and take ownership. I am willing to go back and make quick adjustments (iterating and taking ownership/lead), instead of waiting for someone else to tell me what to do next.

Notice what I did NOT do: I did not try to solve anything yet. I spent the class session understanding the document's structure, its assumptions, and its logic. The Context phase is about building a mental model of the problem space before you touch solutions.

Your Turn — 3 steps to get started

Read the guideline (pages 1-8) with pen in hand — not on a screen. Mark one section where you think a clinical worker would struggle to act on it in the moment. Don’t try to solve it. Just notice where the guideline fails the person reading it.
Write down what you already know and what you don’t know about this domain. Be honest about the gap. What do you know about hospitals, infection control, Bhutan’s healthcare system? What are you guessing? The list of what you don’t know is more useful than the list of what you do.
Complete the Problem Stake Brief — one specific moment, one specific person, one specific constraint. Don’t wait for it to feel ready. Stake a claim, then use the Dojo to pressure-test it. You can iterate — but you can’t iterate on something that doesn’t exist yet.

View Problem Stake Brief → | View Problem Stake Defense →

Four sessions, one arc: Part I is research and planning with AI. Part II is taking that plan into implementation and having AI challenge it. Part III is what happens when you feel overwhelmed and need to simplify. Part IV is iterating on the implementation — pushing back, cross-referencing sources, and building a shareable artifact. Read them in order — each builds on the previous.

Video

AI session: View the ChatGPT conversation →

Handwritten Reflection

Planning: From Stake to Human Process

Open Handwritten PDF →

Covers: what Piranavan changed about my assumptions, picking a specific person, and designing the human process before touching technology.

Takeaways

Write before you prompt. Your thinking shapes the AI session — not the other way around. I wrote ~200 words of my own direction before AI did anything. The push/pull hypothesis, the resources angle, the “high impact with small correction” framing — all of that came from reflection, not from a chat window.
Ask for options, not answers. “Give me three options with pros and cons” — not “tell me the best one.” I kept the decision with myself. That’s the difference between using AI as a tool and outsourcing your judgment to it.
Read everything with pen in hand. I required actual links so I could check sources myself. Then I read the entire response and took notes. If you skip this step, you’re building on a foundation you haven’t verified.

More from this session

Load the session with everything you already have — I attached the guideline, the JSX artifact from the Map phase, copy-pasted the assignment framing, and added my push/pull hypothesis. Don’t start a cold chat.
UMPIRE phases aren’t a rigid sequence — I was in Plan but found myself doing Review and Evaluate on what AI produced. That’s normal. The framework is a thinking tool, not a checklist.
Other people’s perspectives change your thinking — talking to students surfaced the resources angle I hadn’t considered. More conversations = more inputs = better plan.
Catch yourself jumping to solutions too early — I noticed myself sketching color-coded paper clips and screen notices, then pulled back. Noticing that impulse is the skill.
This takes real time — 30 to 45 minutes per phase, not 10. The video shows the setup, not the full work.

Your Turn — 3 steps to get started

Write ~200 words of your own thinking before opening any AI tool. What’s your hypothesis? What direction do you think the solution should go? Load that into AI as your starting point — not a blank prompt asking “what should I do?”
Ask AI for 3 options with pros and cons — then YOU pick. Don’t ask “what’s the best approach?” Ask “give me three options.” Read them, evaluate, and make the choice yourself. The decision stays with you.
Read AI’s full response and take notes by hand. Don’t skim. Don’t copy-paste. Read every paragraph, check any links or sources it cites, and write down what you notice. If you skip this step, you’re building on a foundation you haven’t verified.

View Problem Stake Defense Assignment →

Video

Plan + Implement: From Stake to Build

AI session: View the Gemini conversation →

Handwritten Reflection

From Plan to Implementation

Open Handwritten PDF →

Covers: what changed between the plan and the build, how expert feedback reshaped the approach, and the gap between intention and execution.

Takeaways

Start with your own ideas, then let AI break them. I came in with a specific plan — green/yellow/red patient categorization, wristbands, color-coded clips on paperwork, QR codes to a web app. I didn’t ask AI to generate ideas. I asked it to evaluate mine. It found flaws I hadn’t considered. That’s the value.
AI caught what my engineering mindset missed. Five red patients but only one isolation room — what happens? Color-coded wristbands conflicting with existing hospital color systems. Patients panicking when they see a red tag. Family members who need guidance too. I was solving a technical problem. The AI saw the human problem.
The tool doesn’t matter — the approach does. I switched from ChatGPT to Gemini deliberately. Same method: load context, share your thinking, ask for challenge. The AI is interchangeable. Your process isn’t.

More from this session

Load your context every time you switch tools — I added the guideline, the JSX artifact from Map, and my handwritten notes into a new Gemini gem before writing a single prompt
Ask “do you think this is feasible?” not “build this” — that invites pushback instead of compliance
Low-tech sometimes beats high-tech — a decision tree printed on the back of an intake form might work better than a web app with a QR code
Your earlier UMPIRE work should show up in your plan — I had thought about resource constraints during Map but completely dropped them when I started planning the solution. AI caught the gap I created
There is always more work than you think — the video is 12 minutes, the actual work continues off-camera

Your Turn — 3 steps to get started

Bring your plan to AI and ask it to find the flaws — don’t ask it to build for you. Share your specific plan and ask: “What’s wrong with this? What am I missing?” The value is in what AI breaks, not what it generates.
When AI finds a problem, ask yourself: is this a technical problem or a human problem? Five red patients but one isolation room is logistics. Patients panicking when they see a red tag is human experience. Your solution needs to address both — and you’re the one who sees the human side.
Try a different AI tool with the same method. Load your context, share your thinking, ask for challenge. If your process only works with one tool, it’s not really your process. The approach is yours; the tool is interchangeable.

View Build v1 Assignment →

Video

Implement: Building, Testing, Iterating

AI session: View the ChatGPT conversation →

Handwritten Reflection

Implementation: What Changed and Why

Open Handwritten PDF →

Covers: what changed from the plan to the build, what broke, and what the iteration revealed about the original assumptions.

Takeaways

When you feel overwhelmed, simplify — don’t push through. After the previous session, my stomach dropped. AI threw too many challenges at once and I lost my footing. The fix wasn’t working harder — it was stepping back, reflecting on paper, and narrowing to one decision at a time. I told AI to assume full resources for now. Constraints can be added later. Simplify first, iterate to complexity.
Direct the AI — don’t let it direct you. I explicitly told ChatGPT: keep it conversational, don’t give me pages of response. When AI puts too many options on the table without you making choices, it takes control of the conversation. In the Plan and Implement phases, making choices is YOUR job. If you feel overwhelmed, that’s often a sign AI is making decisions you should be making.
Push back when something feels wrong. “Isolate Now” was technically correct but felt alarming — it would stigmatize patients and scare families. I pushed back. We iterated to “Immediate MDRO Precautions.” The AI accepted it because my reasoning about human impact was valid. Your judgment about how people experience a solution matters as much as technical correctness.

More from this session

Load context from prior tools — I copy-pasted the entire Gemini conversation into ChatGPT so nothing was lost when I switched
One design decision at a time — I spent 15+ minutes just deciding what to call the three categories. That’s not slow, that’s thorough. Rushing past naming means every downstream decision inherits a bad foundation.
Design for action, not description — “What should staff do now?” beats “What kind of patient is this?” Action-first labels reduce translation errors under pressure. This applies to any solution: tell the user what to do, not what category they’re in.
Always cross-reference against primary sources — AI cites the guideline confidently, but you should have the guideline open and checking. I didn’t show that in the video, but you should do it.
Reflection between sessions changes everything — Saturday’s overwhelm led to Sunday’s paper reflection led to Monday’s clarity. The overnight gap wasn’t wasted time. It was processing time.

Your Turn — 3 steps to get started

If you’re stuck, step away and write on paper: what is the ONE decision in front of you right now? Not the whole project. One decision. Name it, make it, then move to the next. Tell AI to assume full resources for now — constraints can be added later. Simplify first, iterate to complexity.
Tell AI what you’ve decided — don’t present what you’re unsure about. Start your next message with “I’ve decided X, here’s why.” When AI puts too many options on the table without you choosing, it takes control. If you feel overwhelmed, that’s often a sign AI is making decisions you should be making.
Find one thing in AI’s response that feels technically right but wrong for your user — and push back. “Isolate Now” was accurate but would stigmatize patients. Your judgment about how people experience a solution matters as much as technical correctness. If you can’t find anything to push back on, you’re not engaging your own judgment.

One test: after your next AI session, can you point to a specific decision YOU made that shaped the direction? If the answer is no, you let AI drive.

Video

Implement: Iterating on the Admission Plan

Conversation Arc

How This AI Session Unfolded

This is a summary of the ChatGPT conversation shown in the video — what I said, what the AI said back, and why each turn mattered.

Loaded context from a prior AI session and presented the Green / Yellow / Red admission tagging plan.
AI pushed back on three structural flaws — overlapping categories, vague “Yellow” criteria, and missing workflow steps.
Simplified to admission-only focus and told the AI to keep responses to 3–5 sentences — it was being too verbose.
AI offered three labeling approaches; chose action-based labels (e.g., “Standard Admission”) over color codes or risk tiers.
Pushed back on “Isolate Now” as a label — too alarming for staff. Iterated to “Immediate MDRO Precautions.”
AI recommended against wristbands; disagreed because wristbands matter when electronic systems fail. Converged on wristbands as a backup signal, not a primary identifier.
Cross-referenced the actual clinical guidelines (Section 2.1) to verify screening criteria before continuing.
Decided to build a shareable artifact for stakeholder review — moving from planning into implementation.

Artifact

Interactive Admission Plan — Built for Stakeholder Feedback

This JSX artifact presents the admission tagging plan, reasoning, and assumptions in a click-through format designed for sharing with the customer and getting feedback from someone with medical background.

Open Interactive Artifact →

Download the JSX for your own work: mdro_admission_plan_review.jsx

Takeaways

Push back, contribute, converge. Don’t accept the AI’s first answer. Challenge recommendations you disagree with (like the wristband exclusion), bring your own reasoning (wristbands matter when systems fail), listen to the AI’s logic (it shouldn’t be the only identifier), and converge on a stronger solution through back-and-forth. This is symbiotic thinking — not delegation.
Cross-reference AI output against primary sources. Before continuing the conversation, the actual clinical guidelines were read carefully to understand screening criteria. AI suggestions were compared against what the guidelines specify — not treated as a substitute for reading them.
Know what decisions you can’t make — flag them for the right person. The screening criteria involved medical assumptions about case categories and device types. Rather than guessing, the decision was explicitly deferred to someone with medical expertise. Build the artifact, but mark the assumptions that need expert validation.

More from this session

Shape the AI’s behavior when it’s not working for you. The AI was too verbose — repeating itself across long paragraphs. Direct instructions were given to keep responses to 3–5 sentences. Manage the tool; don’t let the tool manage you.
Build something shareable to move forward. Instead of staying in the planning loop, the move was to build an interactive artifact presenting the plan, reasoning, and assumptions — something a stakeholder can click through and react to. Implementation creates feedback surface area that planning alone doesn’t.

Your Turn — 3 steps to get started

When AI recommends against something, ask why — then argue back with your reasoning. Don’t accept or reject blindly. State your case, listen to the counter, and converge. The best solutions come from the back-and-forth, not from either side alone.
Open the primary source alongside the AI chat. Whatever guidelines, specs, or documentation your solution is based on — have them open and cross-reference. AI will cite them confidently. Your job is to verify.
Identify at least one assumption in your plan that requires someone else’s expertise to validate. Don’t guess. Mark it, build around it, and take it to the stakeholder. That’s what the artifact is for.

View Solution Brief Assignment →