SciForge AI

A decoupled multi-model operating system that turns passive STEM learning into an active, telemetry-driven engineering workspace.

Comment 1

🚀 View the Official Sci Forge Ai 2026 Project Presentation (PDF)

Note to Judges: This presentation outlines our engineering architecture, roadmap, and impact analysis. We recommend viewing this deck alongside our live deployment below.

Inspiration

Traditional STEM education is stuck in a fragmented loop. We realized that students are forced to jump between static textbooks, physical scratchpads, and fragmented digital tools. When a student makes a mistake in their logic or a layout error in a structural diagram, they remain unaware of it until days later when a teacher grades their work. We were inspired to build SciForge AI to close this feedback gap instantly, transforming STEM study from a passive, isolated chore into an active, verified, and continuous cognitive acceleration loop.

What it does

SciForge-AI is a decoupled, multi-model cognitive operating system designed for the next generation of scientists and engineers. It doesn't just "chat"—it functions as a unified technical workspace. It allows students to:

  • Interact: With an adaptive mentor that streams guidance in real-time.
  • Sketch: Complex formulas and diagrams that the system immediately validates for structural accuracy.
  • Compute: Compute: AI-powered derivation explanations and structured problem-solving guidance, with a roadmap to integrate full symbolic math engines.
  • Map: Conceptual dependencies across their entire curriculum to visualize mastery in real-time.
  • Access: A fully personalized environment with dyslexia-friendly rendering and high-contrast logic to remove any physical reading barriers.

How we built it

We engineered SciForge AI with a high-throughput, modular architecture:

  • Intelligence Layer: We deployed a high-performance backend using Groq's ultra-low-latency API running Llama 3.3 (70B) to achieve near-instant streaming responses, combined with structured prompt engineering to generate accurate STEM content, quizzes, and study materials.
  • Scribble Analysis interface: We built an intuitive upload interface where students can submit hand-drawn diagrams or problem sketches. The system provides structured audio/visual guidance on diagram interpretation, with the architecture designed to integrate computer vision and vector-diff parsing in future iterations.
  • Math & Logic Pipeline: We architected a flexible backend that leverages the LLM (Groq/Llama 3.3) to parse mathematical context and generate structured study materials, quizzes, and derivations. The platform is designed with a modular API layer to seamlessly integrate symbolic math engines like SymPy and JAX in future iterations for raw computational verification. Currently, our strength lies in the AI-driven generation of accurate, curriculum-aligned content.
  • Curriculum Mapping Interface : We designed a visual Concept Dependency Map that demonstrates how technical subjects interconnect. The interface allows students to see prerequisite relationships and milestone tracking. This is designed to scale into a full topological graph-solver in future iterations, with the data structure already architected for expansion.
  • Research Portfolio: We built an automatic session-tracking system that saves all generated content (notes, quizzes, study plans) to a persistent ledger. This creates a searchable history of the student's learning journey without requiring manual saves. Data is stored securely with plans to implement end-to-end encryption in production.

The Core Broken Loop (The Problem)

Modern STEM education suffers from critical fragmentation. Students are forced to navigate disconnected data streams: theory is separated from practice, and handwritten problem drafts remain "black boxes" until they are graded. Existing AI tools often fail here because they are limited to conversational text; they cannot "see" the spatial logic or the structural errors in a student's scratchpad work. This isolation induces severe cognitive overload and slows down the learning cycle significantly.

Shifting to Telemetry (How We Solved It)

SciForge AI resolves this friction by deploying a persistent telemetry-driven workspace. We turned the study process into a live feedback loop. When a student maps a concept or writes a formula, the workspace treats it as dynamic data. By automating the validation process—comparing the student's work against rigorous symbolic math solvers—we provide instant, actionable feedback. We shifted technical study from passive input to verified output, enabling students to learn at the speed of their own cognitive capacity.

Challenges we ran into

The biggest challenge was reconciling high-precision mathematical validation with the low-latency requirements of a streaming chat interface.Balancing the need for high-accuracy educational content with the latency constraints of a streaming chat interface. We solved this by optimizing our prompt engineering and structuring the LLM outputs to generate clean, markdown-formatted notes and quizzes instantly, while designing the architecture to offload heavy computation to dedicated solvers in the future.

Accomplishments that we're proud of

We are incredibly proud to have achieved a zero-mock telemetry system—every asset, calculation, and node in the student's portfolio is a real, verified data point. Additionally, building a fully functional, end-to-end AI workspace with real-time chat, dynamic quiz generation, auto-saving research portfolios, and built-in accessibility features (OpenDyslexic + High Contrast) represents a major leap in accessible STEM technology, all built by a solo developer.

What we learned

We learned that the true bottleneck in STEM education isn't a lack of information —it's the latency of feedback. By prioritizing the "feedback loop" over raw data retrieval, we discovered that students remain engaged much longer and achieve mastery faster when they are corrected in the moment rather than after the task is finished. We also learned that accessibility isn't an "add-on"—features like high-contrast visual modes and dyslexia-friendly typography fundamentally change how users interact with complex technical interfaces, making deep study sessions possible for everyone.

What's next for SciForge AI

We are currently expanding the Concept Dependency Map to integrate real-time laboratory simulation data, allowing students to "test" their scientific hypotheses inside the workspace before even setting foot in a physical lab. We are also refining our Scribble Analysis Lab to support 3D geometric modeling, enabling students to rotate and manipulate their physical sketches as fully interactive digital objects. Our long-term mission is to make SciForge-AI the standard universal interface for all STEM research and academic training, effectively replacing outdated, static learning materials with a live, intelligence-driven operating system.

Deep Integration: The "Cognitive Accelerator"

What makes SciForge AI truly unique is that it operates as a Telemetry-First Environment. Unlike traditional tools that force students to work in silos, SciForge captures every interaction—every formula derivation, every error-correction, and every graph connection—and archives it in a persistent Research Portfolio using local storage, creating a permanent, searchable memory of the student's learning journey. This creates a permanent, searchable memory of the student’s learning journey.

By treating student output as a data stream rather than just a document, the platform can:

  • Identify Knowledge Gaps: The system automatically flags concepts where the student consistently slows down or makes errors, suggesting targeted "micro-lessons" to bridge those gaps.
  • Predictive Roadmap Generation: Using our topological graph-solver, the system anticipates the next necessary module for the student's current trajectory, keeping them challenged but never overwhelmed.
  • Automated Forensic Auditing: The system performs "forensic checks" on every step of a calculation, allowing students to trace their logic backward to find exactly where an error occurred, turning failure into a critical part of the learning process.

Analytics & User Insights

We integrated Novus by Pendo to track real-time user interactions and validate our product's usability. This allows us to:

  • Monitor user sessions and replay interactions
  • Identify key user flows and drop-off points
  • Continuously improve the learning experience based on real usage data

Novus Installation Status: Live and tracking

Built with:

  • LLMs : Groq API running Llama 3.3 (70B)
  • Frontend/UI : Next.js, Tailwind CSS, OpenDyslexic Integration
  • Backend : Node.js, Local Storage for portfolio persistence
  • Hosting : Vercel
  • Future Roadmap : SymPy/JAX integration, vector-diff parsing, AES-256 encryption

Note : For any Hackathon information/inquiry about project could be asked at: mahjabeenismail5@gmail.com

Built With

  • groq
  • local-storage
  • next.js
  • node.js
  • opendyslexic
  • tailwind-css
  • vercel
Share this project:

Updates