AI & ML final year project ideas
Free AI and machine learning final year project ideas on Sefism for CS and software engineering students, each scoped for a single semester with a suggested stack and deliverables.
- 300 project ideas
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.
46 ideas · page 1 of 2
Multi-Agent Literature Review Assistant with Verified Citations for Students
Students writing a literature review often get chatbot answers full of papers that do not exist. Build a multi-agent assistant that plans a search, queries open scholarly databases, reads abstracts and open-access full texts, drafts a structured review, and refuses to cite anything it cannot resolve to a real DOI or arXiv record with a supporting passage. Then measure whether multiple agents actually beat a single agent on quality and cost.
Which LLM Understands Roman Urdu and Bangla Business Messages? A Three-Model Evaluation Harness
Small businesses in Pakistan and Bangladesh get orders and complaints as WhatsApp-style messages in Roman Urdu, Urdu script, Bangla and English mixed together. Build an evaluation harness that compares three LLMs (for example one hosted frontier model, one cheaper hosted model and one open model) on real-shaped tasks such as extracting order details and classifying complaints, with a human-checked gold set, cost per correct answer and a public results page.
Model Card and Datasheet Generator from Training Runs
Model cards and dataset datasheets are the standard way to document what a model does, what data it learned from and where it fails, but they are written by hand at the end, if at all. Build a generator that pulls facts from MLflow runs, dataset files and evaluation results, drafts a model card and datasheet, flags missing or unsupported sections, and keeps the card in sync as the model is retrained.
Tool Selection Evals and Description Linter for MCP Servers
An MCP server can work perfectly and still fail in practice because models pick the wrong tool or fill its arguments badly. Build a tool that lints MCP tool names, descriptions and schemas against good-practice rules, generates realistic user requests for each tool, measures how often several models choose the right tool with valid arguments, and suggests rewrites that improve the score.
MCP Server for Public Transport Timetables Built on GTFS Feeds
Build a Model Context Protocol server that lets any AI assistant answer real transport questions correctly: next departures, journeys with transfers, service changes and accessibility, all computed from GTFS timetable data rather than guessed by the model. Where a city has no public feed, the team builds one for a route they know, such as a university shuttle or a metro bus line, and publishes it openly.
Arabic Diacritics Reading Tutor with Automatic Tashkeel Feedback for Learners
Learners of Arabic, from schoolchildren in Pakistan and Nigeria to expatriate workers in the Gulf, struggle to read text without short vowel marks (tashkeel). Build a tutor that removes the marks from a passage, asks the learner to add them, and checks the answer against an automatic diacritiser and a verified reference, explaining errors by grammar rule. Measure diacritiser accuracy and whether learners improve.
Human-Approved AI Agent for Filing and Tracking Citizen Service Requests
Filing a complaint or service request with a city or government office means finding the right portal, filling long forms, attaching documents and then chasing it for weeks. Build an agent that turns a citizen's plain description (in Urdu, Roman Urdu or English) into a correctly filled request, pauses for the citizen to approve every submission, then tracks status and reminds them to follow up. Built and tested against a realistic mock portal you create.
Drone Plant Stand Counting and Gap Mapping for Smallholder Maize and Sugarcane Fields
A few weeks after sowing, farmers need to know how many plants came up and where the gaps are, so they can resow in time. Build a pipeline that takes low-altitude RGB drone photos of a small field, stitches or tiles them, counts plants with a detection or density model, and produces a gap map and an estimate of plant population per acre. Compare it against hand counts in marked sample plots.
River Flood Forecasting with LSTM Models from Upstream Gauges and Satellite Rainfall
Train and evaluate machine learning models that forecast river discharge one to five days ahead for a few South Asian or African river basins, using upstream gauge readings, satellite rainfall and public hydrology datasets. Compare against simple persistence and published global forecasts, and report the lead time at which a warning threshold would have been crossed in past floods. This is the forecasting model, not a sensor network or an alert app.
Protocol-Locked Danger Sign Triage Assistant for Community Health Workers in Urdu, Bangla and Arabic
Build a phone assistant that walks a community health worker through a sick child visit in Urdu, Bangla or Arabic, where the triage decision comes from an encoded clinical protocol, never from a language model. The model only does what it is good at: understanding a spoken or typed answer in a local language and turning it into a protocol field. The project tests whether that split keeps the decisions safe while making the tool easier to use than a long paper form.
Citation-Checked Legal Question Answering over Pakistan Code and India Code
Law students and junior lawyers in Pakistan and India increasingly ask AI tools legal questions, and those tools sometimes cite sections that do not exist or misquote them. Build a retrieval-augmented assistant over official statute text from Pakistan Code or India Code where every answer cites specific sections, and a separate checker verifies that each cited section exists and that quoted text matches the source before the answer is shown. Measure hallucinated citations before and after the checker.
Bangla Writing Feedback Tutor with Spelling and Grammar Error Detection for School Students
Students in Bangladesh and West Bengal write essays and letters in Bangla for exams, but teachers with large classes cannot give detailed feedback on every piece, and mainstream writing assistants offer little or no Bangla support. Build a tutor that marks likely spelling and grammar errors in a student's Bangla text, explains each one in simple Bangla and tracks a student's common mistakes over time. Evaluate it on real student writing marked by teachers.
Learned LLM Router That Sends Easy Tasks to Cheap Models, with a Cost-Quality Frontier
Many LLM requests are easy enough for a small, cheap model, and a few need a strong, expensive one. Build a router that learns from your own evaluation data which requests need which model, compare it with simple rules and with always using one model, and plot the cost-quality frontier so a team can pick an operating point. Include cost tracking per request and a fallback when the cheap answer looks weak.
Sentence-Level Groundedness Checker for RAG Answers with Human-Calibrated Thresholds
RAG apps cite sources, but a citation does not prove the sentence next to it is supported. Build a drop-in checker that splits each answer into claims, checks every claim against the retrieved passages with a natural language inference model and an LLM judge, highlights unsupported sentences in the UI, and sets its thresholds from human labels rather than guesses. Test it on English and Urdu answers.
Pairwise Preference Labelling App for Urdu LLM Answers with Rater Calibration
Language models learn what 'good' means from human preference data, and very little of it is in Urdu. Build a mobile-friendly app where raters compare two model answers to an Urdu or Roman Urdu prompt, choose the better one against a clear rubric, and explain why. The system checks rater consistency, corrects for position bias, and produces a clean, documented preference dataset and a small leaderboard.
Road Damage Mapping from Phone Video for Municipal Maintenance Teams
Build a system that turns phone video recorded from a car, rickshaw or motorbike into a map of potholes and cracks, merges repeated sightings of the same damage across trips, and gives municipal engineers a ranked repair list. Road surveys in most South Asian towns are done by eye and on paper, if at all. The project fine-tunes an open road damage model on local roads and tackles the harder problem of turning detections into a reliable map.
Mule Account Detection with Graph Machine Learning for Mobile Wallets
Build a simulator of a mobile wallet payment network with realistic scam and mule account patterns, then train and compare graph machine learning models that flag accounts used to collect and move stolen money. Scam proceeds in Pakistan often pass through chains of wallet accounts opened with borrowed or bought identities. Rules catch known patterns; the project tests whether learning from the shape of the money flow catches more, and explains each flag to an analyst.
Medicine Box and Strip Identification for Pharmacy Dispensing Double-Checks
Build a camera check for small pharmacies: the assistant scans the medicine picked from the shelf, and the app matches the box or blister strip to a catalogue, reads the strength and warns when it does not match the prescription line. Look-alike packaging and similar names cause dispensing mistakes, and small pharmacies rarely have barcode systems. The project uses image retrieval with a pretrained model, so new medicines can be added from a few photos without retraining.
Rooftop Solar Panel Mapping from Drone and Aerial Imagery for Electricity Distribution Planners
Build a pipeline that finds rooftop solar panels in drone or aerial images of a neighbourhood, estimates the panel area and a capacity range per building, and totals it by street or transformer area. Rooftop solar has spread fast in Pakistan, much of it outside any register, and distribution engineers cannot see where midday generation will push power back into the network. The project fine-tunes a segmentation model on public solar panel datasets and tests it on local roofs.
Weakly Supervised Urdu and Roman Urdu Classification of Relief Needs During Floods
Build a pipeline that sorts Urdu, Roman Urdu and English messages and posts during a flood into needs such as rescue, food, shelter, medical help and missing people, with place names extracted, so relief coordinators can see requests by area. Instead of labelling thousands of messages by hand during a disaster, coordinators write short labelling rules that a weak supervision framework combines into training data within hours. The project tests how fast a useful classifier can be built this way.
Clustered Short-Answer Grading for Large University Classes with Lecturer Control
Build a grading tool that groups similar typed short answers from an online quiz, lets the lecturer grade and comment on each group once, and flags unusual answers for individual reading. Lecturers with hundreds of students often switch to multiple choice because written answers take too long to mark. The project uses sentence embeddings and clustering, keeps every grade under the lecturer's control and measures how much marking time it saves.
Cloned Voice Detection for Urdu Phone Calls and Voice Notes
Build a detector that tells real Urdu speech from AI-cloned speech in phone-quality audio and compressed voice notes, and test whether it still works on cloning tools it has never seen. Scammers now copy a relative's or a bank officer's voice from a few seconds of audio. Public anti-spoofing datasets are mostly English studio recordings, so the project builds a consented Urdu dataset and measures what really happens after a phone line and a messaging app have squeezed the audio.
Glossary-Consistent English to Urdu Translation for Public Notices and Health Leaflets
Build a translation tool for departments, hospitals and NGOs that translates English notices and leaflets into Urdu while forcing approved terms from their glossary to be used every time, and flags sentences where it could not. Machine translation is fluent but inconsistent: the same medical or legal term comes out three different ways in one leaflet, and some come out wrong. The project compares constrained decoding with an open model, glossary prompting with a language model and a commercial glossary feature.
Pheromone Trap Moth Counting from Phone Photos for Cotton Pest Scouts
Build an app and model that count bollworm moths on a pheromone trap's sticky card from a phone photo, log the count with the trap's location and date, and show agriculture extension staff where catches are rising. Scouts count moths by eye today and write numbers on paper, so the data arrives late and is hard to compare. The project uses an existing phone and trap, not new hardware.
Browse project ideas by domain and difficulty
Each of these is its own page, so you can send someone straight to the AI ideas or the beginner ones.
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.
- 1:1 Mentorship
- Fresh Graduate Mentorship
- Career Guidance
Brand deals, promos and partnerships: tauseeffayyaz@gmail.com
Follow along
New writing, resources and project ideas land here first.