v2.0 · Open source · MIT

Research papers,
turned into
running code.

ScholarDevClaw reads ML research, maps improvements to your codebase, and generates validated patches — autonomously. Runs entirely on your machine.

Active 1260+ tests 6 languages 15 paper specs
01
ingest paper.pdf
Extract algorithms, equations, pseudocode
PDF
02
analyze ./my-repo
AST traversal, framework detection
AST
03
search arXiv 2.4M+
Semantic match to your patterns
2.4M
04
map → generate
6-tier matching + CST patch synthesis
94%
05
benchmark + validate
Real subprocess benchmarks, 1260+ tests
✓ pass
06
open pull request
Clean diff, changelog, confidence scores
PR
RMSNorm· FlashAttention· RoPE embeddings· SwiGLU activations· Grouped Query Attention· LoRA fine-tuning· Speculative Decoding· ALiBi attention bias· RMSNorm· FlashAttention· RoPE embeddings· SwiGLU activations· Grouped Query Attention· LoRA fine-tuning· Speculative Decoding· ALiBi attention bias·

Six phases.
One pipeline.

From raw research paper to production pull request — entirely autonomous, with approval gates where it matters.

01
📄
Analyze Repository

Tree-sitter extracts a full AST from your codebase across 6+ languages. Detects frameworks, patterns, and architecture.

tree-sitter
02
🔍
Search Research

Searches arXiv (2.4M+ papers), Papers with Code, and GitHub for ML research relevant to your specific codebase.

arXiv · PWC
03
🗺
Map to Your Code

6-tier matching: exact name → fuzzy → imports → text scan → legacy → LLM semantic. Every match is confidence-scored.

6-tier
04
⚙️
Generate Patches

15 CST templates (RMSNorm, FlashAttention, RoPE…) plus LLM synthesis fallback for novel improvements.

15 templates
05
📊
Benchmark & Validate

Real subprocess benchmarks measure actual speedup and memory impact. Approval gates on low-confidence changes.

1260+ tests
06
🔀
Open a Safe PR

Clean pull request on a safe branch with full changelog, benchmark results, and confidence scores attached.

git · CI

Built for research
engineers.

Production-grade tooling for teams who operationalize ML research at scale.

  • Multi-language AST — Python, JS/TS, Go, Rust, Java. tree-sitter powered, sub-second analysis.
  • Smart 6-tier matching — never misses an improvement opportunity. LLM semantic search as last resort.
  • Sandboxed validation — all patches run in isolation. 1260+ tests catch regressions before your CI does.
  • Interactive TUI — workflow wizard, live logs, run history, artifact viewer. Full terminal UI.
  • Zero data sharing — runs 100% locally. Your code never leaves your machine.
// example session
# 1. Install
pip install scholardevclaw

# 2. Point at your project
scholardevclaw analyze ./my-transformer
# → Found 6 improvement candidates
# → RMSNorm:    94% confidence
# → FlashAttn:  97% confidence

# 3. Map paper spec → code
scholardevclaw map . rmsnorm

# 4. Generate + validate patch
scholardevclaw generate . rmsnorm
# → Generating patch artifacts...
# → Running 1260 tests... PASS
# → Speedup: +18.4%

# 5. Or just launch the TUI
scholardevclaw tui

Precision at scale.

15 Paper Specs
1260+ Tests Passing
6 Languages
2.4M arXiv Papers Indexed

Running in under
a minute.

Four install methods. Pick whichever fits your environment.

curl -fsSL https://ronak-iiitd.github.io/ScholarDevClaw/install.sh | bash

Common
questions.

ScholarDevClaw is an autonomous Research-to-Code AI Agent. It analyzes your codebase, finds relevant improvements from the latest ML papers, maps them to your code, and generates validated patches automatically. It supports Python, JS/TS, Go, Rust, and Java — running entirely on your local machine.
Generic AI coding assistants complete code you write. ScholarDevClaw proactively reads actual research papers, extracts implementation specs, and autonomously maps them to your specific codebase using 6-tier matching. It's built for researchers and engineers who want to operationalize ML innovations — not just autocomplete.
No. ScholarDevClaw works out of the box with 15 built-in paper specs — RMSNorm, FlashAttention, RoPE, SwiGLU, and more. For advanced LLM-powered semantic matching you can optionally connect Claude, GPT, Gemini, or any other provider.
Never. ScholarDevClaw runs entirely locally. Your code is analyzed on your machine and never transmitted anywhere unless you explicitly configure an external LLM service. Full control, always.
Yes. ScholarDevClaw is fully open source under the MIT license. Fork it, contribute to it, or build on top of it. Source code, docs, and contribution guide are all on GitHub.

Your codebase deserves
the latest research.

Install ScholarDevClaw and run your first research-to-code pipeline in under a minute.