Skip to main content

CodeSeva – AI-Assisted Code Completion for Low-Level Languages

Project Overview#

CodeSeva is an AI-powered coding assistant designed to enhance productivity in low-level programming languages such as 8051, 8085, and ARM assembly.
Unlike traditional code completion tools (focused on high-level languages), CodeSeva bridges the gap between AI and hardware-aware development by integrating fine-tuned LLMs with statistical models for real-time, context-aware suggestions.

[What makes it unique?]
  • Combines Open Mistral 7B (semantic completions), SantaCoder / CodeT5 (error detection & repair), and KenLM N-Gram (Kneser-Ney smoothing) for hybrid predictions.
  • Provides real-time code suggestions, hardware-specific optimizations, and intelligent error detection in a seamless web-based editor.
  • Supports both cloud-powered LLM inference and offline code completion with local N-gram models.

Key Features#

  • Real-time Code Completion

    • Context-aware assembly code predictions tailored for each target architecture.
    • Low-latency hybrid engine: statistical models handle 65–70% of predictions, LLMs manage complex optimizations.

  • Error Detection & Correction

    • Automatically detects syntax, logical, and hardware-related errors.
    • Provides correction suggestions with explanations for better learning.

  • Hardware-Aware Optimizations

    • 8051 → Efficient register bank switching, bit-addressable memory usage.
    • 8085 → Segment register management & string operation optimization.
    • ARM → Pipeline scheduling, conditional execution, and Thumb-2 optimizations.

  • Web Editor with LSP Integration

    • Interactive assembly editor with real-time predictions & debugging.
    • Cross-platform, lightweight, and non-intrusive UI that adapts to user workflow.

  • Continuous Feedback Loop

    • Learns from user interactions (accept/reject suggestions).
    • Implements privacy-preserving retraining for weekly updates.
    • Results in ~0.5% weekly performance improvement.

Performance Highlights#

  • Code Completion:

    • Avg. Accuracy: 80.7%
    • Avg. BLEU Score: 0.82
    • Avg. Code Similarity: 0.81

  • Error Detection (SantaCoder & CodeT5):

    • Token Accuracy: ~92%
    • Exact Match Accuracy: ~79%
    • Response Time: 350–450ms

  • Offline N-Gram Model:

    • Response Time: <10ms
    • Memory: 150MB
    • Accuracy: ~72% of LLM performance
    • Ideal for low-latency or offline coding.
[Developer Impact]

Early adopters reported:

  • 35% reduction in debugging time
  • 28% fewer keystrokes
  • Significant boost in repetitive assembly tasks & embedded workflows

Technical Stack#

  • Models: Open Mistral 7B (quantized), SantaCoder, CodeT5
  • Statistical Engine: KenLM with Kneser-Ney smoothing
  • Frameworks: HuggingFace Transformers, PEFT, Bitsandbytes
  • Backend: Flask + Traefik, REST API integration
  • Data Infra: Elasticsearch, ClickHouse, Airflow
  • Editor: Web-based IDE with LSP support (VS Code integration)
  • Languages: Python 3.11 (core), Cython, Rust (performance-critical paths)

Research Contributions#

[Bridging a Research Gap]

Most AI-assisted code completion focuses on high-level languages (Python, Java, C++). CodeSeva targets assembly and embedded systems, making it one of the first hybrid systems for hardware-constrained environments.

  • Introduces architecture-aware AI coding assistant for legacy + modern microcontrollers.
  • Novel hybrid design: statistical N-gram + transformer-based LLMs.
  • Implements hierarchical attention for semantic understanding of assembly instructions.
  • Privacy-first feedback loop with differential privacy for safe user telemetry.
  • Demonstrates strong cross-platform performance consistency.

Future Enhancements#

  • Natural Language → Assembly Translation
    Write: “Toggle LED on Port 1.0 with 2s delay” → Get ready-to-run 8051 code.

  • Virtual Hardware Simulator Integration
    Run and debug generated assembly in real-time without physical hardware.

  • Intelligent Anomaly Detection
    Identify unusual or inefficient coding patterns before deployment.

  • Multilingual & Educational Support
    Enable global adoption with localized assembly tutorials and auto-explanations.

[Challenges Ahead]
  • Data scarcity for quality assembly training datasets.
  • Variations across hardware vendors (8051/ARM variants).
  • Privacy concerns in handling proprietary low-level code.

Final Thoughts#

CodeSeva is more than just a code completion tool—it’s a step toward the future of AI-assisted embedded systems development. By intelligently blending statistical modeling, LLMs, and compiler techniques, it makes low-level programming:

  • Faster
  • Smarter
  • More Accessible
[Takeaway]

For portfolio readers:
If you’ve ever struggled with assembly language or embedded systems coding, CodeSeva is the kind of AI assistant you wish existed years ago.

Thesis Report

You can download the full CodeSeva Project Thesis here:
📥 Download PDF