Employment Contract and Offer Letter Clause Explainer for Job Seekers
A final year project that reads an employment contract or offer letter, finds clauses such as notice period, probation, non-compete, training bond and IP assignment, and explains each one in plain English and Urdu with the exact source text. You fine-tune a small model for clause extraction and measure it against an LLM baseline.
Intermediate
A small team, or one strong student willing to learn something new.2 semesters · 2 to 3 students
6 to ship
3 optional extrasSuggested stack
What you should ship
- A clause taxonomy of about 15 employment clauses with plain-language definitions reviewed by at least one lawyer or HR professional.
- A labelled corpus of at least 150 employment contracts and offer letters (public templates, published agreements and consented donations) with clause spans marked.
- A span extraction model, first trained on the public CUAD contract dataset and then fine-tuned on your employment corpus.
- A comparison with a prompted LLM on the same test set for span recall, precision, hallucinated clauses and cost.
- A web app where a user uploads a PDF and sees each found clause highlighted in the original, a plain-language explanation, and questions to ask the employer.
- An explicit "not found" list for important clauses that are missing, and a visible statement that the tool gives information, not legal advice.
If you have time left
- Compare two offers side by side on the same clauses.
- Urdu explanations generated from reviewed templates rather than free generation.
- Detect unusual values, such as a notice period or bond amount far above the corpus median.
The problem job seekers have with contracts
Fresh graduates sign employment contracts they do not fully understand. The offer letter is exciting; the twelve-page agreement behind it is not. Clauses that matter a great deal later sit in the middle: a two-year training bond with a large penalty, a three-month notice period during probation, a non-compete that stops you joining a competitor, an IP clause that claims your weekend side project. People discover them only when they try to leave.
Lawyers are expensive for a single contract review, and senior friends are not always available. What a job seeker needs is a clear map of what is in the document, in plain words, with the exact wording so they can ask the right questions before signing.
What already exists and the gap
- Kira and Luminance extract clauses for law firms and corporate legal teams, priced for enterprise due diligence.
- Spellbook helps lawyers draft and review contracts inside Word, aimed at legal professionals.
- General chatbots such as ChatGPT will summarise an uploaded contract, but they may invent clauses, rarely point to the exact text, and give no signal of what is missing.
- Consumer legal AI has a cautionary history: in 2025 the US FTC finalised an order against DoNotPay over "robot lawyer" claims it had not tested.
The gap is a job-seeker tool that extracts employment clauses with evidence, explains them plainly, flags what is missing, and is honest that it is information, not legal advice.
What you build
An upload-and-read app backed by a clause extraction model. The user uploads a PDF. The system extracts text (with OCR fallback for scans), splits it into sections, and runs a span extraction model that finds the start and end of each clause type. For each found clause the app shows the original text highlighted, a short explanation written from a reviewed template, typical values from your corpus, and two or three questions to raise with the employer. A "not found" panel lists important clauses the document does not contain.
The model follows the approach of the CUAD benchmark, which frames contract review as extractive question answering over 41 clause categories in 510 commercial contracts. You start from a model trained on CUAD and adapt it to employment documents with your own labelled corpus.
The technically interesting part
Precision with evidence, and knowing when a clause is absent. Extractive models can only point to text that exists, which limits hallucination, but they still have to decide when to return "no answer". An LLM can paraphrase fluently but may invent a non-compete that is not there. Your core experiment compares the two on three measures: span recall, false clause rate (claiming a clause that is absent), and span overlap with the gold text. Then build a hybrid: the extractive model finds spans, the LLM only rewrites the explanation from the found span. Show your evaluator that the hybrid keeps the low false clause rate of extraction and the readability of generation.
How to evaluate it
- Per-clause precision, recall and token-level F1 for spans on a held-out set of contracts.
- False clause rate for extractive, LLM and hybrid approaches.
- Readability of explanations (a standard readability score) and a correctness check by a lawyer or HR professional on 50 explanations.
- A user test with at least 10 final-year students: after 10 minutes with the tool versus 10 minutes with the raw PDF, can they correctly answer five questions about the contract?
Suggested plan
- Month 1: taxonomy, expert review of definitions, corpus collection.
- Month 2: labelling tool setup and span labelling with agreement checks.
- Month 3: CUAD baseline, fine-tuning on employment data.
- Month 4: LLM baseline, hybrid, false clause experiments.
- Month 5: web app with highlighting, explanations and missing-clause panel.
- Month 6: expert check, user test, write-up.
Path to a product
Offer it free to students through university career offices, which is also where you find test users. Revenue comes from a paid tier for HR teams at small companies who want to check their own templates against common practice, or from licensing the extractor to job platforms that show offers. Other document types (rental agreements, freelance contracts) reuse the same pipeline with new taxonomies.
Scope warning
Do not give legal advice, predict enforceability, or say whether a clause is legal in a given country. Do not store uploaded contracts longer than the session unless the user opts in. Do not try to support every contract type; employment contracts and offer letters are enough.
Learn what you need on Sefism
- AI Engineering roadmap : building and evaluating LLM features with grounded outputs.
- Machine Learning roadmap : fine-tuning transformers and measuring span extraction.
- AI Engineering cheat sheet : patterns for hallucination checks and evaluation sets.
- EU Data Protection Rules for Business : how to think about storing and deleting personal documents.
- Google Colab : free GPUs for fine-tuning on CUAD and your corpus.
Get help with this idea
This project lives or dies on its taxonomy, labelling quality and the honesty of its wording. A mentor can review your clause list and the disclaimer copy before you build the app. Bring your draft scope, the dataset you plan to use and your evaluation table, and we can look at them together. The build is yours; the guidance is on scope, stack and how to measure it. If you are new here, create a free account first.
Ideas and guidance, not finished projects
These are project ideas and scoping guidance, published free for students to use as a starting point. I do not build, write, or sell final-year projects, and I do not complete coursework for anyone. Take an idea, make it yours, and build it.
Not sure this one fits you?
I help students pick an idea that matches their skills and their deadline, then scope it down until it is finishable. Guidance only, never done for you.
Stuck on something specific?
Writing only gets you so far. If you want an answer to your situation rather than the general case, book a session and we will work through it together. Sessions are free for approved Sefism members, and a few slots open each week.
Follow along
New writing, resources and project ideas land here first.