Generative AI & LLM Engineering Internship

Go past prompt tinkering into retrieval, evaluation and cost control.

Most people can get an LLM to produce something plausible. Far fewer can make it reliable, grounded and affordable. This programme is about the second part: retrieval-augmented generation, evaluation you can trust, and the token maths that decides whether a feature ships.

Intermediate4 weeks · remote4 briefs · complete any 2Free to join

What you'll be able to do

Build a RAG pipeline that cites its sources
Evaluate an LLM feature with numbers rather than vibes
Control token cost deliberately
Explain when fine-tuning beats retrieval, and when it does not

Who it suits

Anyone who has used the ChatGPT API and wants to understand what it takes to put an LLM feature in front of real users.

What you'll work with

Python
LangChain
Vector DBs (Chroma/FAISS/Qdrant)
Embeddings
Streamlit/FastAPI

The project briefs

Complete any 2 of these 4to finish the programme. Assessment criteria are published here so you know what “good” means before you start.

Document Q&A with Citations

~12h estimated

Answer questions over a document set, with every answer traceable to a source passage.

Chunk, embed and index the corpus. Retrieve, then generate with the retrieved context. The hard part is chunking — experiment with at least two strategies and report which won.

Assessed against

Two chunking strategies compared with evidence
Every answer shows the passages it used
The system says "I don't know" when retrieval returns nothing relevant
Corpus of at least 50 pages

Suggested datasets

Any public PDF corpus — research papers, government reports, product docs

You'll finish with

GitHub repo
Demo video
A note on which chunking strategy won and why

RAG Evaluation Suite

~10h estimated

Prove your RAG system works, with a number you would defend in a review.

Build a question set with known answers. Measure retrieval quality (did the right passage come back?) separately from generation quality (was the answer right given the passage?). They fail differently.

Assessed against

30+ question/answer pairs
Retrieval and generation scored separately
At least one deliberately unanswerable question
Results as a table, with an interpretation

You'll finish with

Evaluation repo
Results table
Written analysis

LLM Cost Optimisation Study

~8h estimated

Cut the cost of an LLM feature without materially hurting quality.

Take a working feature and attack cost: shorter prompts, cheaper models for easy cases, caching, batching. Measure quality before and after so you can prove what the saving cost you.

Assessed against

Baseline cost per 1,000 requests, measured not estimated
At least three optimisations applied
Quality measured before and after
Final recommendation with the trade-off stated

You'll finish with

Repo + cost analysis
Before/after comparison table

Structured Output Extraction Pipeline

~10h estimated

Turn messy unstructured text into validated structured data, reliably.

Use schema-constrained generation and validate every output. Anything that fails validation must be retried or routed for review — never silently accepted.

Assessed against

A typed schema the output must satisfy
Validation with an explicit failure path
Measured extraction accuracy on 50+ samples
Handles documents that contain none of the target fields

Suggested datasets

Invoices, résumés, job posts, or public filings

You'll finish with

GitHub repo
Accuracy report

Free to apply and free to complete. The completion certificate is optional.

Other domains

AI Agents & Automation

Build autonomous agents that use tools, not chatbots that answer questions.

Machine Learning

Classical ML done properly — including the evaluation everyone skips.

Data Science

Turn a messy dataset into a decision somebody can act on.

Ready to Transform Your Enterprise with AI?

Book a free 30-minute consultation with our AI architects. Discover how agentic AI can automate your workflows and multiply your team's output.

No commitment required · Free architecture review · NDA protected

Generative AI & LLM Engineering Internship

Go past prompt tinkering into retrieval, evaluation and cost control.

Intermediate4 weeks · remote4 briefs · complete any 2Free to join

What you'll be able to do

Build a RAG pipeline that cites its sources

Evaluate an LLM feature with numbers rather than vibes

Control token cost deliberately

Explain when fine-tuning beats retrieval, and when it does not

Who it suits

Anyone who has used the ChatGPT API and wants to understand what it takes to put an LLM feature in front of real users.

The project briefs

Complete any 2 of these 4to finish the programme. Assessment criteria are published here so you know what “good” means before you start.

Document Q&A with Citations

~12h estimated

Answer questions over a document set, with every answer traceable to a source passage.

Chunk, embed and index the corpus. Retrieve, then generate with the retrieved context. The hard part is chunking — experiment with at least two strategies and report which won.

Assessed against

Two chunking strategies compared with evidence
Every answer shows the passages it used
The system says "I don't know" when retrieval returns nothing relevant
Corpus of at least 50 pages

Suggested datasets

Any public PDF corpus — research papers, government reports, product docs

You'll finish with

GitHub repo
Demo video
A note on which chunking strategy won and why

RAG Evaluation Suite

~10h estimated

Prove your RAG system works, with a number you would defend in a review.

Assessed against

30+ question/answer pairs
Retrieval and generation scored separately
At least one deliberately unanswerable question
Results as a table, with an interpretation

You'll finish with

Evaluation repo
Results table
Written analysis

LLM Cost Optimisation Study

~8h estimated

Cut the cost of an LLM feature without materially hurting quality.

Take a working feature and attack cost: shorter prompts, cheaper models for easy cases, caching, batching. Measure quality before and after so you can prove what the saving cost you.

Assessed against

Baseline cost per 1,000 requests, measured not estimated
At least three optimisations applied
Quality measured before and after
Final recommendation with the trade-off stated

You'll finish with

Repo + cost analysis
Before/after comparison table

Structured Output Extraction Pipeline

~10h estimated

Turn messy unstructured text into validated structured data, reliably.

Use schema-constrained generation and validate every output. Anything that fails validation must be retried or routed for review — never silently accepted.

Assessed against

A typed schema the output must satisfy
Validation with an explicit failure path
Measured extraction accuracy on 50+ samples
Handles documents that contain none of the target fields

Suggested datasets

Invoices, résumés, job posts, or public filings

You'll finish with

GitHub repo
Accuracy report

Free to apply and free to complete. The completion certificate is optional.