AI Engineer Intern Evaluation — Imperio Railing Systems

Imperio Railing Systems

AI Engineer Intern — Evaluation

This is the application for our AI Engineer Intern role. We are looking for builders — people who ship working systems, not just slide decks. The form has five sections and should take 60–90 minutes. You can save progress in the browser and come back.

Estimated time: 60–90 min Sections: 5 Auto-saved locally: yes

About You

~3 min

Quick contact details so we can reach you. All fields required unless noted.

Full name*

Email*

Phone (with country code)*

City*

LinkedIn URL*

GitHub URL (strongly preferred)

Portfolio / personal site (optional)

Earliest start date*

Internship duration you can commit*

Hours per week*

Resume

~1 min

Upload your most recent CV. PDF preferred (DOCX accepted). Max 5 MB.

Click to upload or drag and drop

PDF / DOCX, up to 5 MB

AI Knowledge Test

~15 min

10 questions. Answer in your own words — please do not paste from ChatGPT. We are looking for clarity of thinking, not perfection.

Q1 · MULTIPLE CHOICE

What does "RAG" stand for in the context of LLM applications?

Recursive Agent Generation Retrieval-Augmented Generation Reinforced Adversarial Grounding Random Access Generation

Q2 · SHORT ANSWER

In 2–3 sentences, when would you fine-tune a model vs. use a system prompt + RAG? Give a concrete example.

Q3 · MULTIPLE CHOICE

A vector database is best described as:

A relational database optimised for very wide tables A store for high-dimensional embeddings that supports similarity search A graph database with bidirectional edges A queue system for inference requests

Q4 · SHORT ANSWER

Name 3 frameworks/tools you would consider for building an LLM agent, and one strength of each.

Q5 · SHORT ANSWER

An LLM keeps confidently citing a product spec that does not exist in our catalog. What is happening, and what are 2 ways to reduce it?

Q6 · MULTIPLE CHOICE — pick all that apply

Which of the following are reasonable use-cases for embeddings? (Select all that apply.)

Semantic search over a product catalog De-duplicating customer support tickets Storing exact transaction dates for audit Finding similar past projects to a new inquiry Calculating monthly payroll totals

Q7 · SHORT ANSWER

Briefly explain "tool use" / "function calling" in LLMs and one place you would use it for a sales workflow.

Q8 · SHORT ANSWER

Your prompt is producing inconsistent JSON. What are 3 specific things you would try, in order?

Q9 · MULTIPLE CHOICE

You're spending too much on tokens for an internal Q&A bot over 1,200 product PDFs. The fastest cost win is usually:

Switching from a hosted model to running an open-source model on a GPU box Adding more system-prompt instructions to keep answers shorter Better retrieval — fewer, more relevant chunks instead of stuffing context Caching every full conversation in memory

Q10 · SHORT ANSWER

Name one production "gotcha" with LLM apps that you have personally hit (or read about and would actively design around). Be specific.

Business Case Studies

~25 min

Three short cases grounded in our actual business. We are looking for: (a) clear problem framing, (b) concrete stack choices, (c) prioritisation, (d) awareness of how this lands with non-technical users.

CASE 1 · SALES

The cold-lead problem

Imperio receives roughly 30 product inquiries a day across email, the website form, and WhatsApp. Sales reps reply to the first 5–10 and the rest go cold within a week. Many of those cold leads are residential customers who would have bought if we had followed up well.

Design an AI-driven follow-up system. Address:

How leads are captured and classified
What the AI does autonomously vs. what a human approves
Stack you would use (be specific — model, orchestration, channel)
What you would measure to know it is working

CASE 2 · ERP / QUOTING

AI-assisted quote generation

A new inquiry comes in: "Need 50m of L-Series aluminium railing for a 4th-floor open balcony, frost finish, project in Pune, install by end-July." Today a sales rep manually pulls product details, checks stock, applies pricing rules, and emails a PDF quote. This takes 40–90 minutes and is the #1 bottleneck before close.

Walk through how an AI agent would generate this quote. Address:

How the agent extracts structured fields from the inquiry
How it accesses the product catalog and pricing rules
Where you would put a human-in-the-loop
What can go wrong, and how you would catch it before the quote goes to the customer

CASE 3 · POST-SALE

Turning installations into intelligence

Most of our customer relationship dies after install. We do not know which products cause issues, which contractors do clean work, or which customers would refer us. Design an AI-driven post-installation workflow that turns finished projects into useful data.

Address:

Timing and channel for the touchpoints (be specific)
What data you would collect, and what AI does with it
Two examples of decisions this data could improve

Execution & Build Challenge

~30–45 min, plus build time

This section matters most. We hire builders. Talk less about what you would do and more about what you have done.

Link to your most impressive AI project*

Repo, demo, or live URL. If private, add a short Loom walkthrough below.

What problem did it solve, in one sentence?*

What % of the code did you personally write?*

Loom or video walkthrough (strongly preferred — 3 to 5 min)

Share your screen, walk through the code and the system. We weight this highly.

One thing about this project you would do differently if you started over*

Build Challenge — optional but heavily weighted

Build a small AI agent that reads a one-page product PDF (we provide a sample railing brochure at the link below) and answers natural-language questions about it via a chat UI or CLI. Constraints:

Must use a real LLM API (any provider) and a retrieval step over the PDF
Spend no more than 4 hours on it
Ship a public repo with a README explaining how to run it
Record a 2–4 minute Loom showing it work end-to-end and walking through the code

Sample PDF: we will email it after submission. Or use any one-page product PDF you have access to and tell us which.

Build challenge — repo URL (optional but heavily weighted)

Build challenge — Loom URL

Build challenge — total time spent

Self-assessment (1 = never used, 5 = built production systems with it)

LangChain / LangGraph

1 2 3 4 5

RAG / vector search

1 2 3 4 5

Prompt engineering

1 2 3 4 5

Python / backend

1 2 3 4 5

Workflow tools (n8n / Make / Zapier)

1 2 3 4 5

Shipping things end-to-end

1 2 3 4 5

In one sentence: why you, for this role?*

How did you find out about this role?

By submitting, you confirm the work above is your own and that any code linked is yours or appropriately credited. We will get back to you within 7 working days.

Auto-saving locally…

Imperio Railing Systems · AI Engineer Intern Hiring · Confidential

Application received

Thanks for the time you put into this. We will review and reply within 7 working days. If you submitted the build challenge, we will weight it heavily.