Template preview — Draft structure. Links, metrics, certificates, and LeetCode stats are placeholders — nothing here is claimed as real.

MD IrfanFull-stack developer
Résumé
All work

Case study 02 — Draft

WeatherGPT

A draft case study for WeatherGPT. Add the real stack, the prompt strategy, and one outcome. Nothing below is a claim yet.

Role
Add role
Timeline
Add timeline
Stack
Add stack
Links
Add live URLAdd repo

Illustrative mockup — replace with a real screenshot.

The short version

  • What it is — add one sentence.
  • What you built — add the model, the forecast source, and the flow.
  • What was hard — add the grounding or latency problem.
—Add one outcomea number you can defend

The finished story. Switch to Process for the sketches, dead ends, and the bug that ate a week.

Who it was for

To add

Add the real user and the question they wanted answered. “Will it rain at 7 pm?” is a better starting point than “weather enthusiasts”.

The problem

To add

Add the gap between a weather dashboard and a plain answer. Note the risk: a language model can write a confident sentence about weather it does not have.

How it works

To add

Add the real flow: geocoding, forecast source, model, streaming, cache, and rate limit. Name the exact boundary where facts stop and prose begins.

01Clientadd the framework and the screen
02Serveradd the API, auth, and cache
03Dataadd the database or the external API
Draft illustrative architecture — replace the nodes with the real stack
This diagram is a placeholder. It shows the shape of the explanation, not the stack.

The hard part

To add

Add the grounding or evaluation problem. What did the model invent, how did you detect it, and what did you change? Link the prompt or the eval when it exists.

The outcome

To add

Add one number you can defend: first-token time, cache hit rate, or eval pass rate. If there is no number, describe the guardrail instead.

What I would change

To add

Add one honest thing: parse relative times on the server, cache data instead of prose, or put the eval suite in CI.

Draft case study. The structure is real; the details are placeholders. The visitor-weather demo in the mobile case study uses real browser data; this page does not yet have a live demo.

The process

Three moments from the build. Add real dates and commits when they exist; until then, this is the shape of the story.

  1. Step 01

    The first version

    Add what the first version did and how it failed. If it hallucinated, say so; that is the most useful part of the story.

  2. Step 02

    Facts before prose

    Add how the forecast reaches the model: structured context, a tool call, or a guardrail. Explain what the model is not allowed to do.

  3. Step 03

    What broke

    Add the failure that survived the first fix: stale data, time parsing, a cost spike, or an answer with no source.