SmartCourse: AI-Driven Academic Advising

• SmartCourse is an integrated advising system that uses AI-driven analysis of transcripts and degree plans to generate personalized course recommendations.
• The system fuses administrative data with a locally hosted large language model, ensuring context-aware suggestions that enhance curricular alignment and GPA recovery.
• Evaluated using custom metrics such as PlanScore and PersonalScore, SmartCourse demonstrates improved personalization compared to traditional, generic advising tools.

SmartCourse is an integrated course management and AI-driven advising system for undergraduate students, specifically tailored to the Computer Science (CPS) major, that combines transcript and plan information with a locally hosted LLM for personalized course recommendations (Mi et al., 26 Jul 2025). It was designed to address the limitations of traditional advising tools, particularly their tendency to provide generic advice and to remain context blind with respect to a student’s transcript and official degree plan. The system combines administrative functions, student-facing academic workflows, and transcript-aware recommendation generation, and it is evaluated through custom advising metrics—PlanScore, PersonalScore, Lift, and Recall—defined to measure curricular alignment and personalization under different context conditions (Mi et al., 26 Jul 2025).

1. Advising problem and design rationale

SmartCourse emerges from a specific diagnosis of undergraduate advising. Existing degree-audit or dashboard systems can show completed credits or unmet requirements, but they rarely “think” like an advisor; they issue boilerplate reminders without tailoring suggestions to the student’s unique history. Prior systems, as described in the SmartCourse paper, do not fuse a student’s transcript with the official degree plan, and therefore cannot distinguish between courses already completed, courses in which the student struggled, or electives that best serve the student’s interests (Mi et al., 26 Jul 2025).

The design target is the Computer Science undergraduate curriculum. CS majors must juggle core requirements such as programming, data structures, and theory alongside specialized tracks such as AI, cybersecurity, and software engineering. SmartCourse is explicitly motivated by advising queries of the form “Which electives will build toward a machine-learning specialization?” and “Should I retake that algorithms course to boost my GPA?” It therefore treats degree completion, GPA recovery, prerequisite sequencing, and elective selection as part of a single advising problem rather than as separate tools or dashboards (Mi et al., 26 Jul 2025).

In the broader course-recommendation literature, other systems have emphasized different objectives. A web-based recommender for the Virtual University of Pakistan uses user-based collaborative filtering and rating-prediction to estimate expected marks and grades (Akhtar, 2020), while a goal-based recurrent model recommends preparatory courses for a target course by modeling prior knowledge and the Zone of Proximal Development (Jiang et al., 2018). SmartCourse is distinctive in centering transcript-aware natural-language advising grounded simultaneously in academic record and official program plan (Mi et al., 26 Jul 2025).

2. System architecture and academic operations

SmartCourse follows a modular, layered design. Its user-facing components are split between a CLI and a Gradio web GUI. The CLI is used by system administrators to manage the course catalog, user accounts, and LLM configuration, including switching models via Ollama. The Gradio GUI provides a chat-style interface for instructors and students: instructors can record grades, review transcripts, and audit student progress, while students can register or drop courses, view their four-year plan, and request AI-based advising (Mi et al., 26 Jul 2025).

The core academic modules comprise user account management, course enrollment and grading, and degree plan management. User account management stores hashed credentials, roles such as student, instructor, and admin, and declared major(s) in a secure account store implemented as a flat file. A “course-catalog” repository holds all course codes and titles, while an “enrollment ledger” tracks each student’s enrollments and grades. For each major—initially CPS—a four-year plan file lists the 39 required courses by year, and a progress monitor compares the plan against the transcript to compute unmet requirements (Mi et al., 26 Jul 2025).

The recommendation subsystem is coupled to these academic records rather than to an isolated recommendation index. SmartCourse uses a locally hosted LLM, specifically llama3.1:8b, deployed via the Ollama runtime. An SMTP-based mailer sends emails for new recommendations, enrollment confirmations, or grade postings. This yields a unified platform in which advising is not external to academic operations but embedded in account, enrollment, grading, and plan-management workflows (Mi et al., 26 Jul 2025).

This architecture differs from systems that operate primarily on catalog text. For example, a separate LLM-based course recommender uses a Retrieval Augmented Generation pipeline in which GPT-3.5-turbo generates an “ideal” course description, embeddings are compared against a corpus of precomputed course-description vectors, and GPT-4o selects final recommendations from retrieved context (Deventer et al., 2024). SmartCourse instead grounds generation in institutional student data—transcript and degree plan—rather than in catalog semantics alone (Mi et al., 26 Jul 2025).

3. Contextual prompting and recommendation logic

On each advising request, SmartCourse gathers three inputs: transcript data, outstanding plan requirements, and the student’s natural-language question. These are concatenated into a single structured prompt and passed to the LLM over a Python subprocess. The paper gives the prompt form as:

“Transcript: [Course1: Grade, Course2: Grade, …] Remaining Requirements: [CourseA, CourseB, …] Question: [Student’s question text]”

The system then parses the LLM’s text response into a recommendation set $\mathcal{R}$, filters out any non-catalog courses, and returns the final suggestions to the user (Mi et al., 26 Jul 2025).

The contextual logic is twofold. By including the transcript, the model knows which prerequisites have been satisfied and which courses carry low grades, defined in the paper as below $B{-}$. This enables recommendations to retake low-grade courses and to avoid recommending courses already passed. By including the degree plan, the model is grounded in official requirements and does not stray into irrelevant electives. The sample prompt template further constrains the output with the instruction “Only recommend courses from the official catalog. Explain your reasoning in 2–3 sentences per suggestion” (Mi et al., 26 Jul 2025).

Post-processing is deliberately simple. After receiving the LLM’s output, SmartCourse applies a simple parser to extract course codes; any course not found in the course-catalog repository is dropped. This filtering stage is presented as a mitigation for hallucinations rather than as a substitute for model grounding (Mi et al., 26 Jul 2025).

In methodological terms, this transcript-and-plan conditioning contrasts with semantic retrieval systems that optimize representation geometry over course descriptions. One recent system fine-tunes BERT with contrastive loss and isotropy regularization and returns Top-$N$ courses by cosine similarity in an embedding space (Khreis et al., 16 Jan 2026). SmartCourse’s advising mechanism is not an embedding-retrieval pipeline; it is a contextual prompting workflow whose grounding variables are institutional transcript and plan data (Mi et al., 26 Jul 2025).

4. Evaluation framework and custom metrics

SmartCourse was evaluated on 25 representative advising queries, including “Suggest electives for an AI specialization,” “Courses to improve GPA with low-grade retakes,” and “Electives for a Ph.D. track in ML.” For each query, the system was run in four context modes: Full Context (transcript + plan + question), No Transcript (plan + question), No Plan (transcript + question), and Question-Only (question alone) (Mi et al., 26 Jul 2025).

The evaluation introduces four custom metrics. Let $\mathcal{P}$ denote the set of outstanding plan requirements, $\mathcal{L}$ the set of low-grade courses, and $\mathcal{R}$ the set of courses recommended by the LLM. Then:

$$\mathrm{PlanScore} = \frac{\lvert \mathcal{R}\cap\mathcal{P}\rvert}{\lvert\mathcal{R}\rvert}$$

$$\mathrm{PersonalScore} = \frac{\lvert \mathcal{R}\cap(\mathcal{P}\cup\mathcal{L})\rvert}{\lvert\mathcal{R}\rvert}$$

$$\mathrm{Lift} = \mathrm{PersonalScore}-\mathrm{PlanScore}$$

$$\mathrm{Recall} = \frac{\lvert \mathcal{R}\cap\mathcal{P}\rvert}{\lvert\mathcal{P}\rvert}$$

These metrics are tailored to advising rather than to pure ranking. PlanScore measures the fraction of recommendations that are still required, PersonalScore adds retake advice, Lift quantifies the gain from personalization through retakes, and Recall measures coverage of remaining requirements (Mi et al., 26 Jul 2025).

This choice of metrics is notable because the surrounding literature often uses substantially different evaluation criteria. Collaborative-filtering course recommenders have been assessed with Mean Absolute Error (Akhtar, 2020), semantic retrieval systems with Hit-Rate@5, F1@5, MRR@5, and IsoScore (Khreis et al., 16 Jan 2026), and hybrid skill-aware recommenders with precision@5, recall@5, and nDCG@5 (Soni et al., 12 Nov 2025). SmartCourse’s metric family is therefore specific to the problem of contextual academic advising rather than generic item recommendation (Mi et al., 26 Jul 2025).

5. Empirical findings

The core empirical result is that using full context yields substantially more relevant recommendations than context-omitted modes. The averages reported over 25 queries are as follows (Mi et al., 26 Jul 2025).

Mode	Context	Key averages
Full Context	transcript + plan + question	6.6 recs; PlanScore 0.53; PersonalScore 0.78; Lift 0.25; Recall 0.15; Latency 47.7 s
No Plan	transcript + question	2.2 recs; PlanScore 0.03; PersonalScore 0.19; Lift 0.16; Recall 0.01
No Transcript	plan + question	6.2 recs; PlanScore 0.60; PersonalScore 0.69; Lift 0.09; Recall 0.17
Question-Only	question alone	0.04 recs; PlanScore 0.04; PersonalScore 0.04; Lift 0.00; Recall 0.00

The ablations isolate the role of each context component. Omitting the degree plan virtually destroys PlanScore and Recall because the model no longer knows what remains to be taken. Omitting the transcript yields high PlanScore because the model simply parrots outstanding requirements, but personalization drops sharply, with Lift falling to 0.09 from 0.25. In Question-Only mode, the LLM produces almost no valid suggestions. Statistical significance was confirmed via bootstrap 95% confidence intervals (Mi et al., 26 Jul 2025).

The qualitative analysis is consistent with the quantitative results. Full-context advice was described as coherent, respecting prerequisites and suggesting retakes of low-grade courses. Hallucinations—defined here as courses outside the official catalog—occurred in less than 10% of outputs and were easily caught by the catalog filter. The paper concludes that full context nearly doubles both curricular alignment and personalization compared to context-free modes and that transcript-aware advice can suggest retakes and avoid redundant course recommendations (Mi et al., 26 Jul 2025).

6. Limitations, extensions, and research context

The paper presents SmartCourse as a prototype rather than a production-complete advising platform. Its current constraints are explicit: evaluation was performed on only one CPS student and 25 synthetic queries; hallucinations and bias can persist even with filtering; the prototype does not yet enforce FERPA/GDPR encryption or access controls; the PlanScore/PersonalScore family treats the plan as ground truth and ignores valid electives outside that plan; and the approximately 48-second latency per full-context query is slow for interactive use (Mi et al., 26 Jul 2025). A plausible implication is that institutional deployment would require both security hardening and substantial systems optimization.

The proposed extensions are correspondingly concrete. These include multi-major support through additional plan repositories, user feedback loops such as like/dislike signals, enriched context using course descriptions, student career goals, and extracurriculars, a full web portal beyond Gradio with schedule planners and real-time dashboards, and performance tuning through prompt caching, model distillation, or asynchronous pre-fetch (Mi et al., 26 Jul 2025).

Within the wider literature, these directions align SmartCourse with several adjacent research trajectories. One line integrates career interests, resumes, transcripts, and job-posting data into a skills-aware hybrid recommender with content-based filtering, collaborative filtering, and matrix factorization (Soni et al., 12 Nov 2025). Another integrates course and job data through a heterogeneous graph and skill-community detection to support cross-domain career-education recommendation (Zhu et al., 2020). Yet another studies natural-language course discovery via RAG over course descriptions (Deventer et al., 2024). SmartCourse occupies a more tightly institutional position: it is a contextual advising system grounded in transcript and degree-plan data, coupled directly to course management operations, and evaluated by metrics intended to capture not only curricular compliance but also retake-aware personalization (Mi et al., 26 Jul 2025).