Course guide

College Project Tracks

Applied engineering and research directions for 3rd- and final-year CSE students

This guide turns the weekend's small prototypes into realistic college projects. The course gives you a starting artifact; it does not make a production system in two days. It also covers the part before the code: where ideas actually come from, how to pick one, how to form a team that survives the semester, and how to build with AI assistants without renting your own project.

The rule that keeps a project gradeable

Choose one user, one data source, one decision, and one measurable claim.

Bad scope:

Build an AI assistant for students.

Good scope:

Help the exam-cell assistant answer 40 recurring questions from the current regulations PDF, with citations, at least 85% answer accuracy on a 100-question test set, and a refusal when the answer is not in the document.

Scope check · tick what is true of your idea
Four ticks: gradeable, start. Three: fix the blank before you write code. Two or fewer: that is "an AI assistant for students" wearing a different shirt.

Where ideas actually come from

Good ideas are collected, not brainstormed. A 1 a.m. hostel brainstorm produces "AI assistant for students"; an hour of watching the exam cell produces a project. Go looking in this order, because the closer the source is to you, the better your data access, and data access decides projects.

On campus: problems with a face attached

Off campus: problems with data attached

One warning: an idea that arrives with no reachable user and no accessible data is not an idea. It is a tweet.

How to pick one: five tests, two minutes each

Run every candidate through these five before you commit anything.

Idea filter · one candidate idea, five honest ticks
Five ticks: start this week. Four: fix the gap first. Three or fewer: kill it tonight. Killing an idea in week 1 costs one evening; killing it in week 8 costs the semester.

And pick the problem, not the stack. Nobody grades you for using an agent framework. You are graded for a number.

The idea wall: twenty problems worth hacking on

These are seeds, not scopes. Each still has to pass the five tests, the gradeable rule, and ship the nine artifacts below. Steal freely, then narrow.

Idea dealer · one seed at a time

Mess menu truth-meter

Photograph the day's mess board, extract the dishes, diff against the published weekly menu.

Field accuracy on 20 board photos, with a rejection path for unreadable boards

Exam-cell queue killer

RAG over the current regulations PDF answering the 40 questions the exam cell hears daily, with citations and refusal.

85% on a 100-question set that includes unanswerable questions

Bus-route decoder

Bilingual Q&A over the city and college bus charts, Tamil and English.

Accuracy on 30 real student questions asked in both languages

Lost-and-found matcher

Embed found-item photos and lost-item descriptions, match across the modality gap.

Top-3 hit rate on a 25-pair test set

Timetable clash oracle

Q&A over the master timetable: free rooms, staff clashes, lab overlaps.

30 clash questions with zero false "room is free" answers

Placement JD mapper

Extract skills from job descriptions and cite which subject and unit covers each.

Extraction accuracy plus mapping agreement with a placement coordinator

Circular digest

Bilingual college circulars to one-line summaries; dates and fees flagged, never paraphrased.

Faithfulness on 40 circulars and zero missed deadlines

Scholarship form pre-checker

Vision check of a filled form against its instruction page, listing what will get it rejected.

Rejection-reason recall on 20 filled forms

Lab-record rubric checker

Scanned lab record against the rubric: missing aim, missing output, unsigned pages.

Section-detection agreement with the lab in-charge

PYQ topic miner

Cluster five years of question papers by topic, citing question numbers per topic.

Topic-label agreement with a subject senior on 100 questions

Quiz forge

Unit-wise quizzes generated from your own notes with page citations.

Valid-question rate judged blind by a senior

Attendance oracle

"How many periods can I still miss?" answered only from the regulations, refusing when they are silent.

Zero answers without a cited clause

Symposium helpdesk

Event FAQ bot where the schedule changes hourly and answers carry timestamps.

Stale-answer rate across three schedule updates

Library shelf-finder

"Where is this book, and is it in?" over a catalogue export.

Top-1 accuracy on 30 lookups including misspelled titles

Hostel complaint router

Classify and route complaints to electrical, plumbing, wifi, or mess, with human approval before dispatch.

Routing accuracy plus a blocked-action log for every auto-dispatch attempt

Ledger digitizer

A local shop's handwritten ledger, often Tamil, to structured rows.

Character error rate per field, refusal on illegible rows

Four-language city guide

Multilingual RAG over official tourist information: Tamil, English, Hindi and one more.

Language-wise answer accuracy and token-cost ratio per language

Department style-cop

Check student C or Python code against the department coding-standard sheet.

Violation precision and recall on 20 already-graded submissions

Lecture-transcript QA

One course's recorded-lecture transcripts, answers with timestamp citations.

Citation-correctness rate on 30 questions

Notice-board digest

A weekly photo of the physical notice board into a deadline-sorted digest.

Extraction accuracy and missed-deadline count over one month

Every project must ship these artifacts

  1. Problem brief: named users, current workflow, and the specific pain.
  2. Data note: source, ownership, privacy classification, and update frequency.
  3. Architecture diagram: model, prompt, retrieval, tools, interface, and storage.
  4. Baseline: the current process or a simpler model/prompt that your approach must beat.
  5. Evaluation set: 20 examples for a small prototype, 50 for a semester project, or 100 for a final-year claim. Include unanswerable and failure cases.
  6. Failure report: at least three failures, their category, and the mitigation attempted.
  7. Cost and latency note: tokens, requests, approximate cost, p50 latency, and what happens on timeout or quota exhaustion.
  8. Safe demo: no secrets in the repository, no private data in free-tier prompts, no destructive action without human approval.
  9. README and reproducible run: setup, sample data, test command, limitations, and a short demo video or hosted URL where appropriate.

A polished UI without these artifacts is a demo. A modest UI with these artifacts is an engineering project.

Choose a track

Be honest about two things: the hours your whole team will truly spend, and what must exist at the end. Match them to a level; ambition one level above your hours does not label test sets.

Track picker · two honest questions
Hours the whole team will truly spend
What must exist at the end
Answer both. Honesty is cheaper here than at review 2.

Level 1 · 6–10 hours after the course

These are small portfolio extensions.

ProjectCore techniquesMinimum evidence
Study-notes RAGembeddings, retrieval, citations, evaluation20 questions, one baseline, one retrieval failure
Receipt/document extractorvision, structured output, validationfield-level accuracy on 20 images or synthetic samples; unknown path for unreadable fields
Prompt evaluation dashboardprompt versioning, regression tests, latency/cost loggingcompare two prompts on a fixed test set and display failures

Level 2 · 12–20 hours

These fit a mini-project or a guided lab series.

ProjectCore techniquesMinimum evidence
Department regulations assistantRAG, citations, refusal, freshness, injection testing50 labelled questions, document dates, answer and retrieval metrics, poisoned-document test
Codebase RAG assistantcode chunking, semantic + keyword retrieval, citationsindex one student repository, answer 30 code questions, cite file paths and line ranges
College document QAmultimodal extraction, schema validation, human reviewmarksheet/form/receipt pipeline, field-level confusion table, explicit rejection cases
Subject tutor for one courseRAG, quiz generation, evaluation, optional visionone subject only, 30-question test set, source citations, weak-topic report
Tamil or multilingual scheme assistantmultilingual retrieval, token measurement, RAG, evaluationparallel language questions, token/cost comparison, language-specific failure analysis

Level 3 · 20–40 hours

These are suitable for a semester project when one team owns the scope.

ProjectCore techniquesMinimum evidence
LLM evaluation and observability dashboardevals, traces, cost, latency, prompt versionsrun history, p50/p99 latency, token/cost charts, regression alerts, privacy-safe logs
RAG + tools assistant with approval gatesRAG, function calling, workflow, validation, human approvalat least two read-only tools, one blocked side effect, step cap, tool-choice evaluation
Department knowledge serviceingestion pipeline, metadata filters, hybrid retrieval, deploymentupdate/re-index path, citations, access control, 100-question evaluation, hosted demo
Personal AI tutor for one subjectmultimodal input, RAG, quiz/eval loop, user feedbackmeasurable learning task, baseline comparison, failure and bias analysis, privacy plan

Form the team before the code

The team you form in week 1 decides more of your grade than the model you call in week 6.

How to form one

How not to form one

Each of these sinks real teams every year.

The sixty-second team audit

Team audit · tick every sentence that is true of your team
Zero ticks: rare, go build. One or two: fix them in the working agreement this week, in writing. Three or more: that is not a team, it is a WhatsApp group. Re-form it while it is still cheap.

A good semester progression

Weeks 1–2 · Narrow the problem

Interview 3–5 users. Observe the current process. Collect only the documents you are allowed to use. Write one measurable claim.

Weeks 3–4 · Build a baseline

Start with a plain prompt or manual workflow. Record where it fails. Do not add RAG, tools, or agents before you have a baseline.

Weeks 5–6 · Build the smallest useful system

Add one technique at a time: structured output, retrieval, or one read-only tool. Keep the system understandable enough to draw on a whiteboard.

Weeks 7–8 · Create the evaluation set

Write labelled examples with a domain expert. Include ambiguous, unanswerable, multilingual, and adversarial cases. Freeze version 1 of the set.

Weeks 9–10 · Measure and harden

Compare against the baseline. Separate retrieval failure from generation failure. Test prompt injection, stale documents, malformed output, timeouts, and quota errors.

Weeks 11–12 · Deploy and explain

Put the key server-side, add limits and logs, deploy a small interface, run a user trial, and document what you would not automate.

Use AI tools like an engineer, not a vending machine

You will build this with AI assistants; the course itself assumes it. The difference between teams that AI accelerates and teams it sinks is a handful of habits, and examiners have learned to test for them.

Three rules that are not optional

  1. The viva rule. Never commit a line you cannot explain at a whiteboard. If AI wrote it and you cannot defend it, you did not build a project; you rented one, and viva is the repossession.
  2. The model never grades its own homework. Your evaluation labels are written and verified by humans. Use a model to draft questions or find edge cases, fine; the correct answers are yours. A test set generated by the system under test is self-deception with a spreadsheet.
  3. The privacy line. Real marksheets, mess bills with names, medical certificates: never into a free-tier prompt. Anonymize or synthesize first. This repeats artifact 8 on purpose, because it is the one that gets colleges into newspapers.

Where AI multiplies you

Where it quietly costs you marks

Habits that compound

Level 4 · Research-oriented final-year projects

These are not “build a chatbot” projects. Each one is an experiment with a hypothesis, a controlled comparison, a measurable result, and a reproducible artifact. A team should choose one track and implement a narrow version over a semester; do not attempt all the techniques listed in a row.

Research trackResearch question / hypothesisBaselines and variablesMetrics and artifact
Multilingual and code-switched RAGDoes retrieval quality degrade when students ask the same question in English, Tamil, Tanglish, and Hindi? Does query translation improve recall without increasing hallucination?Baselines: English-only dense retrieval and keyword search. Variables: language, translation, embedding model, chunking.Recall@k, MRR, answer faithfulness, token/cost ratio, language-wise error report, reproducible multilingual benchmark.
Chunking and hybrid retrieval studyDoes structure-aware chunking plus BM25+dense retrieval beat paragraph-only dense retrieval on college regulations and code?Baselines: fixed-size chunks and dense-only search. Variables: chunk size, overlap, metadata filters, BM25 fusion, reranking.Recall@k, MRR, nDCG, answer accuracy, latency, index size, ablation table, reusable retrieval benchmark.
Hallucination and abstention calibrationCan a system learn when to answer and when to say “I don’t know” without becoming uselessly evasive?Baselines: always-answer prompt and grounded prompt. Variables: refusal threshold, retrieval score, verifier model, answer format.Coverage, selective accuracy, abstention precision/recall, calibration curve, risk-weighted error, refusal/evidence dashboard.
Multimodal document QA robustnessWhich image conditions cause structured extraction to fail: blur, skew, lighting, handwriting, language, or small text? Can preprocessing improve field accuracy?Baselines: raw image at one resolution. Variables: crop, resize, deskew, contrast, prompt, schema, model.Field-level accuracy, character error rate, rejection rate, confidence/error calibration, robustness matrix, labelled image fixture set.
Agent trajectory reliabilityDoes a guarded workflow outperform an unconstrained agent on the same task as tool count and task length increase?Baselines: direct prompt and unguarded loop. Variables: step cap, validation, retries, model size, workflow/agent choice.Task success, per-step success, end-to-end reliability, step distribution, unnecessary calls, cost, failure trace viewer.
Prompt-injection defense evaluationWhich defense layers reduce successful instruction hijacking, and what false positives do they create?Baseline: naive RAG prompt. Variables: delimiters, instruction hierarchy, input filtering, output validation, tool permissions, human gate.Attack success rate, benign-task pass rate, leakage rate, false-positive rate, blast-radius analysis, red-team corpus and report.
Quality–cost–latency model routingCan a router send routine questions to a small/local model and hard questions to a stronger model while preserving quality?Baseline: always use the strongest API model. Variables: routing features, confidence threshold, model pair, caching, output cap.Quality, cost/query, p50/p95 latency, escalation rate, worst-case failures, routing policy and simulation notebook.
Local versus API model under constraintsFor a bounded college task, when does a quantized local model become preferable to an API model?Baselines: API model and a smaller local model. Variables: quantization, context length, language, prompt, hardware.Accuracy/faithfulness, tokens/sec, TTFT, RAM, energy proxy, privacy exposure, cost at 100/10k users, deployment decision.
Evaluation-set quality and judge biasHow much do conclusions change when the evaluator is biased, the test set is contaminated, or the test questions are written by different groups?Baselines: exact/contains scorer and human labels. Variables: LLM judge, position swap, question author, difficulty, language.Human–judge agreement, position bias, inter-rater agreement, score intervals, contamination audit, evaluation methodology paper.

What makes one of these research, not just a bigger app?

A final-year team should submit all of the following:

  1. Research question: one sentence that could be false.
  2. Hypothesis: what you expect and why.
  3. Dataset protocol: how data is collected, labelled, split, versioned, and legally handled.
  4. Baseline: a simple method that a reviewer can reproduce.
  5. Controlled variables: change one major factor at a time; record model/version/prompt/configuration.
  6. Ablation study: remove one component and show what changes.
  7. Metrics: define them before seeing the result; report confidence intervals or repeated-run spread where practical.
  8. Error taxonomy: categories, examples, counts, and representative failures.
  9. Reproducibility pack: code, environment, configuration, fixture data or data-generation script, and exact commands.
  10. Limitations and ethics: privacy, bias, data rights, misuse, and what the system must refuse to automate.

A feasible research-paper structure

SectionWhat students should write
AbstractProblem, method, comparison, main number, limitation.
IntroductionUser/problem context and why the question matters.
Related work5–10 relevant papers or systems, not a list of buzzwords.
MethodData pipeline, architecture, variables, and evaluation protocol.
ExperimentsBaseline, ablations, metrics, repeated runs, and resource budget.
ResultsTables and plots with the failures visible, not only the best number.
Error analysisWhy the system failed and which component was responsible.
Ethics and limitationsPrivacy, fairness, security, generalization, and non-goals.
ReproducibilityRepository, environment, commands, and data statement.

Research complexity ladder

For most final-year teams, Level B or C is the right ambition. Level D can be excellent, but only when the dataset and evaluation are genuinely strong.

What not to choose for this course

Do not begin with autonomous research agents, multi-agent swarms, Kubernetes troubleshooting, GitHub issue-to-PR automation, distributed training, full RLHF, vision-language pretraining, speculative decoding servers, or a generic chatbot that answers everything. These can become later research directions, but they are poor first college projects because the scope and evaluation boundary are unclear.

Suggested evaluation rubric

CriterionWeight
Problem clarity and user evidence15
Working artifact and reproducibility20
Baseline and evaluation quality25
Failure analysis and mitigation15
Security, privacy, and responsible use15
Communication and documentation10

The best project is not the one with the most agents. It is the one that solves a real, bounded problem and can show a number, a baseline, and an honest limitation.

After this guide

Use the current course artifacts as your starting point:

For deeper theory, use the external AI Engineering from Scratch curriculum selectively: classical ML evaluation, self-attention, tokenizer construction, LLM application engineering, tool protocols, and observability. Its full 503-lesson curriculum is a long-term roadmap, not a two-day syllabus.