AnveshanaAI: Adaptive AI Learning Platform

• AnveshanaAI is a multimodal platform that uses a fine-tuned LLM and SBERT-based semantic filtering to generate questions aligned with Bloom’s taxonomy.
• It features a modular microservices architecture with dedicated learner modules like Playground, Challenges, Simulator, Dashboard, and Community for adaptive, interactive assessments.
• The system gamifies progression and employs explainability via heatmaps, enabling both real-time adaptive feedback and educator oversight.

Searching arXiv for the specified paper to ground the article and confirm citation details. AnveshanaAI is a multimodal, application-based learning platform for artificial intelligence that combines automated question generation, adaptive assessment, gamified progression, interactive simulation, and explainable AI within a modular microservices architecture. The platform is designed around learner-facing modules—Playground, Challenges, Simulator, Dashboard, and Community—alongside a parallel administrator interface, and it targets structured learning across domains such as data science, machine learning, deep learning, transformers, generative AI, LLMs, and multimodal AI, with scope to include more in the future (Thakur et al., 28 Sep 2025). Its central technical claim is that an end-to-end system combining generative LLMs, semantic-similarity checks, Bloom’s-taxonomy labeling, adaptive delivery, and explainable attributions can support balanced learning progression and measurable engagement gains beyond static question repositories (Thakur et al., 28 Sep 2025).

1. System architecture and platform organization

AnveshanaAI is built as a fully modular, microservices-driven platform. Its learner-facing surface is divided into five modules—Playground, Challenges, Simulator, Dashboard, and Community—while a parallel administrator interface supports management and oversight (Thakur et al., 28 Sep 2025). The high-level data flow begins when a learner’s HTTP request, such as “fetch next question,” reaches a React/Vite front end styled with Tailwind CSS. The request is then forwarded to a Node.js/Express API gateway, which orchestrates downstream calls to specialized services for question generation, adaptive delivery, simulation, and storage (Thakur et al., 28 Sep 2025).

The back end is organized around multiple Docker-packaged microservices. The Question Generation Pipeline uses a fine-tuned Mistral-7B with LoRA adapters for generation. The Adaptive Engine is a Python service running learner-profiling algorithms. The Simulator provides Jupyter-like notebooks in containers for code and simulation tasks, and the Analytics Dashboard aggregates logs via Kafka streams into MySQL and a reporting service (Thakur et al., 28 Sep 2025). Persistent storage is distributed across MySQL for structured data, FAISS for fast embedding retrieval, and an object store based on S3 for logs and media (Thakur et al., 28 Sep 2025).

The architecture reflects a separation between interaction, orchestration, model inference, and analytics. This suggests a design intended to support low-latency educational workflows while permitting independent evolution of question generation, retrieval, simulation, and dashboarding components. The explicit inclusion of FAISS, SBERT-based retrieval, and a difficulty-estimation module described as C-BERT style indicates that content delivery is not solely generative, but mediated by retrieval and ranking layers that constrain output quality and pedagogical fit (Thakur et al., 28 Sep 2025).

2. Automated question generation and ranking

The core assessment workflow is the Automated Question-Generation Pipeline. It consists of three main stages: retrieval of a seed context from the vector database via SBERT embeddings, optional difficulty-scaling augmentation through paraphrasing or domain shift, and autoregressive generation by a fine-tuned Mistral-7B (Thakur et al., 28 Sep 2025). The paper presents the following pseudocode for the workflow:

$e_q$3

Generation is trained with the standard cross-entropy loss:

$\mathcal{L}_{\mathrm{gen}(\theta) = -\sum_{(c,q)}\sum_{t=1}^{|q|}\log P_\theta\bigl(q_t\mid q_{<t},c\bigr)\,.$

To prevent generated questions from becoming either trivial rewrites or semantically stray items, AnveshanaAI applies semantic-similarity checks between each candidate question $q$ and the source context $c$ using SBERT vectors $e_q$ and $e_c$:

$$\mathrm{sim}(q,c) = \frac{e_q\cdot e_c}{\|e_q\|\;\|e_c\|}\,.$$

Candidates with $\mathrm{sim}\in[0.6,0.8]$ are preferred, with the stated goal of balancing coherence and novelty (Thakur et al., 28 Sep 2025). A margin-rank loss then orders high- versus low-quality items:

$\mathcal{L}_{\mathrm{rank} = \sum_{(q^+,q^-)}\max\bigl(0,\;\delta - \mathrm{sim}(q^+,c)+\mathrm{sim}(q^-,c)\bigr)\,.$

These mechanisms define a hybrid generation regime in which retrieval, augmentation, language-model decoding, and semantic ranking are tightly coupled. A plausible implication is that the system is intended to avoid the two failure modes common in educational item generation: excessive lexical overlap with source material and off-target abstraction. The preference interval for cosine similarity formalizes that trade-off directly rather than treating novelty as an unconstrained by-product of generation (Thakur et al., 28 Sep 2025).

3. Adaptive assessment and Bloom’s taxonomy integration

AnveshanaAI frames assessment as both domain-aware and adaptive. Once questions are generated and presented, real-time performance signals—specifically correct/incorrect outcomes and time-to-answer—are fed back into an adaptive selector (Thakur et al., 28 Sep 2025). The simplified decision rule for difficulty stepping is given as

$d_{n+1} = d_n + \alpha\,(p_{\text{target} - p_n),$

where $p_n$ is the user’s rolling success rate and $q$0 is the target mastery, given in the description as an example of $q$1 (Thakur et al., 28 Sep 2025). Hints are triggered when time exceeds $q$2 or incorrect attempts exceed 2, and hint selection is also guided by Bloom level (Thakur et al., 28 Sep 2025).

Bloom’s taxonomy is integrated at the data-schema level. Every QA entry is tagged with a Bloom level $q$3, and the on-disk schema is specified as

$q$4

Internally, Bloom is treated as an ordinal scale: 1 = Remember, 2 = Understand, 3 = Apply, 4 = Analyze, 5 = Evaluate, and 6 = Create (Thakur et al., 28 Sep 2025). This explicit ordinal treatment is important because it allows the platform to align generation targets, hint selection, and adaptive sequencing with a pedagogical progression rather than only with topical categories or coarse difficulty bins.

The coupling of Bloom labels with adaptive delivery distinguishes AnveshanaAI from static repositories organized primarily by topic or fixed problem sets. The paper states that balanced learning progression is ensured through a dataset grounded in Bloom’s taxonomy, and that adaptive, automated, and domain-aware assessment methods are employed (Thakur et al., 28 Sep 2025). This suggests that progression is intended to be multidimensional, combining domain coverage, item difficulty, and cognitive level rather than relying on a single scalar notion of mastery.

4. Explainable AI and auditability

AnveshanaAI incorporates explainable AI techniques to make the generation process transparent. The platform computes token-level attributions via gradient-based saliency. Given the final token logit $q$5 for a candidate question, the importance of input token $q$6 is defined as

$q$7

These attribution scores are rendered as heatmaps, and a global explanation-score ranks questions according to

$q$8

The stated purpose is to allow educators to audit which tokens drove the LLM’s choice (Thakur et al., 28 Sep 2025). In the platform’s overall framing, explainability is tied not only to model interpretability but also to reliability and transparency in educational content generation.

This use of saliency-derived heatmaps is notable because it positions explainability as an operational feature of the assessment pipeline rather than as an after-the-fact analysis tool. The paper’s discussion states that explainability heatmaps and open-schema labeling engender trust among educators (Thakur et al., 28 Sep 2025). A plausible implication is that the platform treats educator oversight as a necessary complement to automated generation, particularly in settings where question quality, pedagogical appropriateness, and semantic faithfulness must be reviewed rather than assumed.

5. Gamification, learner modeling, and personalization

Gamified tracking is a defining part of the learner experience. The platform presents a personalized dashboard featuring streaks, levels, badges, and structured navigation across supported AI/ML domains, and it incorporates points and achievements to enhance engagement and learning (Thakur et al., 28 Sep 2025). Switching between Playground, Challenges, Simulator, Dashboard, and Community is described as supporting exploration and collaboration (Thakur et al., 28 Sep 2025).

Points accrue per completed challenge according to its difficulty weight $q$9:

$c$0

where $c$1 is a base-point unit, with the example $c$2 given in the description (Thakur et al., 28 Sep 2025). Levels are computed as

$c$3

and day-streaks increment when a challenge is solved within 24 h of the previous one. Badges are thresholded at predefined milestones such as “1000 points,” “7-day streak,” and “Bloom 6 mastery” (Thakur et al., 28 Sep 2025).

Personalization is driven by a learner profile vector $c$4, updated after each session using a recency-weighted rule:

$c$5

Here, $c$6 represents success or failure on a topic, and the resulting vector drives recommended categories on the dashboard by prioritizing areas with low $c$7 (Thakur et al., 28 Sep 2025). This profile update rule is simple, but it operationalizes personalization as ongoing estimation of topical weakness rather than only as retrospective performance logging.

The gamification subsystem therefore has a dual role. On one side it provides observable progression markers—streaks, points, levels, badges. On the other it becomes part of the personalization loop by linking challenge completion, topical outcomes, and dashboard recommendations. The platform’s framing is that adaptivity, gamification, interactivity, and explainability are integrated rather than isolated features (Thakur et al., 28 Sep 2025).

6. Experimental evaluation

The paper reports experiments spanning dataset composition, annotation structure, fine-tuning behavior, and learner engagement (Thakur et al., 28 Sep 2025). The dataset comprises 10,845 QA pairs distributed across 26 subject categories, 4 difficulty levels, and 6 Bloom levels. The reported annotation-quality summary is as follows.

Dimension	Coverage statistics
Subject categories	26 subject categories; effective = 16.57, entropy = 4.051, concentration = 0.044
Difficulty levels	4 difficulty levels; effective = 3.65, entropy = 1.866, concentration = 0.053
Bloom levels	6 Bloom levels; effective = 5.84, entropy = 2.546, concentration = 0.011

Cross-dimensional correlations are reported via Cramér’s V coefficients. Category–Difficulty has $c$8 and is labeled Strong, while Category–Bloom level has $c$9 and Difficulty–Bloom level has $e_q$0, both labeled Weak (Thakur et al., 28 Sep 2025). The strong Category–Difficulty association and weak associations involving Bloom level suggest that Bloom labels are not merely a proxy for content area or difficulty band. A plausible implication is that the dataset supports relatively independent control over cognitive level during question selection and generation.

The fine-tuning configuration uses Mistral-7B v0.1 with 4-bit quantization and LoRA adapters. Reported hyperparameters are batch = 2, LR = $e_q$1, weight_decay = 0.01, epochs = 5, and eval interval = 700 steps (Thakur et al., 28 Sep 2025). Loss curves are described as showing stable convergence by approximately 15k steps. Validation perplexity dropped from 1.5 to 1.3, and the final end-to-end perplexity measured 2.04. Against held-out human questions, BERTScore F1 is 0.427, with precision 0.289 and recall 0.818 (Thakur et al., 28 Sep 2025).

The engagement pilot involved 200 learners over a 4-week period. Reported outcomes were an average streak length increase of 2.3 days with $e_q$2 under a paired t-test, a 12% increase in mean session length, and an 18% increase in completion rate of mid-level Bloom tasks (Thakur et al., 28 Sep 2025). The paper interprets these gains as confirming that adaptive delivery plus gamification sustains engagement. From an evaluative standpoint, the results combine behavioral metrics with generation metrics rather than treating model quality and educational usage as separate concerns.

7. Interpretation, limitations, and prospective extensions

The paper’s discussion identifies several key findings. It states that a synthesized pipeline combining on-the-fly generation, semantic-balance filtering, and domain-aware adaptivity can out-perform static question banks in both coverage and engagement (Thakur et al., 28 Sep 2025). It also states that explainability heatmaps and open-schema labeling engender trust among educators (Thakur et al., 28 Sep 2025). These conclusions position AnveshanaAI as a platform in which pedagogical structure, model generation, and learner telemetry are co-designed.

The limitations are also stated explicitly. Precision of generation remains modest, with BERTScore-Precision = 0.289, and this leads to occasional overly verbose items (Thakur et al., 28 Sep 2025). In addition, the current adaptive logic is rule-based, and the paper notes that deeper personalization may require reinforcement-learning approaches (Thakur et al., 28 Sep 2025). These caveats are significant because they distinguish between demonstrated functionality and still-open optimization problems in item quality and sequencing policy.

Proposed enhancements include integrating peer-review loops in the Community module for human–AI hybrid validation, expanding to new domains such as reinforcement learning and graph ML without manual schema changes, and incorporating curriculum-level planning via a reinforcement learner that sequences Bloom levels automatically (Thakur et al., 28 Sep 2025). These proposals indicate that future work is aimed at two forms of extension: broader domain coverage and more sophisticated policy learning.

A common misconception would be to treat the platform as merely a gamified quiz system. The technical description does not support that reduction. Its defining characteristics are the interaction of generative LLM-based question production, SBERT-based retrieval and semantic filtering, Bloom-tagged schema design, rule-based adaptive delivery, saliency-based explanations, and module-level support for coding and simulation tasks (Thakur et al., 28 Sep 2025). Conversely, it would also be inaccurate to interpret it as a fully autonomous pedagogical agent: the emphasis on explainability, educator audit, and proposed peer-review loops indicates a continuing role for human oversight.