← All projects/Case study № 01

Lumen An AI study companion that shows its sources.

Placeholder. One paragraph summary: the problem, what you built, and the single most important result.

Role
Solo builder — research, design, code
Timeline
8 weeks · 2026
Stack
Python · LangChain · FAISS · Streamlit
Type
Personal project → campus pilot
Cover image placeholder — 16:8
01 — Overview

Why this, why now.

Placeholder paragraph. Set the scene: who the users are, what they were struggling with and why it seemed worth an afternoon — which became eight weeks.

Placeholder paragraph. A note on constraints: time, data, budget, the fact that you were also studying for exams.

02 — The problem

Students didn’t trust AI answers.

Placeholder paragraph. Describe the problem in the user’s words. Include one piece of evidence — a survey, an interview quote, a number.

“It sounds confident, but I can’t tell whether it came from my syllabus or from the internet.”

— Placeholder user interview
03 — Approach

Small loops, real users.

Placeholder paragraph. How you broke the problem down and the order you tackled it in.

  1. Research

    Placeholder — interviews with 12 students.

  2. Prototype

    Placeholder — a notebook version in a weekend.

  3. Test

    Placeholder — five students, one exam week.

  4. Iterate

    Placeholder — citations became the core feature.

04 — What I built

A pipeline you can inspect.

Placeholder paragraph. Explain the architecture in plain language first, then the interesting technical decisions.

PythonLangChainFAISSStreamlitOpenAI API
05 — Results

What changed.

40

Placeholder — students used the pilot during exams.

62%

Placeholder — returned in the second week.

4.6/5

Placeholder — average usefulness rating.

06 — Reflection

What I’d do differently.

Went well

Keep

  • Placeholder — talking to users before writing code.
  • Placeholder — making citations the core feature.
  • Placeholder — shipping something small, early.
Next time

Change

  • Placeholder — define a success metric on day one.
  • Placeholder — evaluate answer quality systematically.
  • Placeholder — think about cost per query sooner.
Next case study

Atlas Retail demand forecasting

Read next