NLP Engineering Internship

Make machines read text well enough to act on it.

Language work did not stop being useful when LLMs arrived — it got cheaper and more measurable. This programme covers the parts that still decide whether a text system works: extraction, classification, evaluation and knowing when a smaller model beats a prompt.

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

What you'll be able to do

Build a text classifier and know why it fails
Extract structured data from messy documents
Evaluate a language system beyond accuracy
Judge when fine-tuning beats prompting on cost and quality

Who it suits

Developers comfortable with Python who want language skills that survive the next model release.

What you'll work with

Python
spaCy
Hugging Face Transformers
scikit-learn
PyTorch
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.

Text Classifier with an Honest Baseline

~10h estimated

Classify real text, and prove your model beats the dumb approach.

Start with TF-IDF and logistic regression. That is your baseline. Then try a transformer. If the transformer does not clearly beat the baseline, that is a finding worth reporting, not a failure to hide.

Assessed against

A stated baseline, with its score
Per-class precision and recall, not just accuracy
A confusion matrix with the worst confusions discussed
An honest note on whether the complex model earned its cost

Suggested datasets

AG News
IMDb reviews
Kaggle consumer complaints

You'll finish with

GitHub repo
Short evaluation write-up

Structured Extraction from Messy Documents

~12h estimated

Pull specific fields out of unstructured text reliably.

Invoices, resumes or job posts. The hard part is not the happy path — it is the document that is missing a field, or has it twice. Decide what your system does then, and say so.

Assessed against

Extracts at least five distinct fields
Returns a confidence or a null, never a guess presented as fact
Evaluated against a hand-labelled set of at least 50 documents
Documented behaviour for missing and duplicate fields

Suggested datasets

Resume datasets on Kaggle
Public invoice samples
Job posting scrapes

You'll finish with

GitHub repo
Labelled evaluation set

Semantic Search over Your Own Corpus

~12h estimated

Build search that finds things keyword search misses.

Embed a corpus, store the vectors, retrieve by similarity. Then measure it: pick 20 queries, note what should come back, and see whether it does. Hybrid keyword + vector usually beats either alone — check whether it does here.

Assessed against

A corpus of 500+ documents
Retrieval quality measured on a fixed query set
Comparison against plain keyword search
A working search interface, even a crude one

You'll finish with

GitHub repo
Query set with results
Live demo or video

Language Model Evaluation Harness

~10h estimated

Build the thing that tells you whether a prompt change made it worse.

A test set, a scoring function, and a runner that compares two versions. This is unglamorous and it is what separates people who ship language features from people who tinker with prompts.

Assessed against

At least 30 test cases with expected behaviour
Automated scoring, not eyeballing
Side-by-side comparison of two prompt or model versions
A documented case where the harness caught a regression

You'll finish with

GitHub repo
Evaluation report comparing two versions

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.

Generative AI & LLMs

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

Machine Learning

Classical ML done properly — including the evaluation everyone skips.

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

NLP Engineering Internship

Make machines read text well enough to act on it.

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

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.

Text Classifier with an Honest Baseline

~10h estimated

Classify real text, and prove your model beats the dumb approach.

Assessed against

A stated baseline, with its score
Per-class precision and recall, not just accuracy
A confusion matrix with the worst confusions discussed
An honest note on whether the complex model earned its cost

Suggested datasets

AG News
IMDb reviews
Kaggle consumer complaints

You'll finish with

GitHub repo
Short evaluation write-up

Structured Extraction from Messy Documents

~12h estimated

Pull specific fields out of unstructured text reliably.

Invoices, resumes or job posts. The hard part is not the happy path — it is the document that is missing a field, or has it twice. Decide what your system does then, and say so.

Assessed against

Extracts at least five distinct fields
Returns a confidence or a null, never a guess presented as fact
Evaluated against a hand-labelled set of at least 50 documents
Documented behaviour for missing and duplicate fields

Suggested datasets

Resume datasets on Kaggle
Public invoice samples
Job posting scrapes

You'll finish with

GitHub repo
Labelled evaluation set

Semantic Search over Your Own Corpus

~12h estimated

Build search that finds things keyword search misses.

Assessed against

A corpus of 500+ documents
Retrieval quality measured on a fixed query set
Comparison against plain keyword search
A working search interface, even a crude one

You'll finish with

GitHub repo
Query set with results
Live demo or video

Language Model Evaluation Harness

~10h estimated

Build the thing that tells you whether a prompt change made it worse.

A test set, a scoring function, and a runner that compares two versions. This is unglamorous and it is what separates people who ship language features from people who tinker with prompts.

Assessed against

At least 30 test cases with expected behaviour
Automated scoring, not eyeballing
Side-by-side comparison of two prompt or model versions
A documented case where the harness caught a regression

You'll finish with

GitHub repo
Evaluation report comparing two versions

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