🧪 Collabuild
An Intelligent Multi‑Agent Framework for Research Paper–to–Prototype Generation
✨ Unique Selling Propositions
- ✅ Rapid Prototyping – 3 hours → under 5 minutes (97.6% time saving)
- ✅ 8 LLM Providers – OpenRouter, NVIDIA, Claude, Ollama, KoboldCPP, Bhashini AI, Dev (offline)
- ✅ Multilingual NLP – Bhashini AI integration for 24 Indian languages
- ✅ Local Deployment – CLI + Docker, works fully offline with Ollama / KoboldCPP
- ✅ Data Attribution – All generated artifacts cite source research papers
- ✅ IEEE 830‑compliant SRS – Structured, consistent, actionable
Investigators: N.S. Sai Karun, K. Gayathri, K. Kishore · PB Siddhartha College, Vijayawada
Guide: Dr. K. Udaya Sri (HoD, DS&AI) · 2026
1 / 16
📄 Abstract
Collabuild automates the transformation of academic research papers into build‑ready software prototypes using nine specialised AI agents. It generates IEEE 830‑compliant SRS, designs architecture, plans SDLC, and produces deployment‑ready code.
Key Results: 97.6% time saving · 78.4% SRS accuracy · 3.9/5 expert quality · 40 papers across 4 domains.
Keywords: Multi‑Agent Systems, Large Language Models, SRS, Research‑to‑Software Translation, AI‑Augmented SE, Bhashini AI, Prototype Generation, SDLC Automation.
📌 USP Highlight: Unlike other frameworks, Collabuild supports local deployment (Ollama/KoboldCPP), multilingual Indian language NLP, and data attribution – every requirement is traced back to the source paper.
2 / 16
❓ Problem & Rationale
Statement of the Problem
- Manual extraction of requirements from research papers is time‑consuming and error‑prone.
- Ad‑hoc system design leads to inconsistent documentation.
- Extended development cycles delay validation of research concepts.
- Existing frameworks lack research‑paper input, multi‑provider flexibility, and Indian language support.
Rationale
- Bridge the academic‑industrial gap.
- Accelerate innovation from weeks to minutes.
- Standardise documentation (IEEE 830).
- Democratise AI with open‑source + free LLM tiers.
- Empower Indian researchers with Bhashini AI (24 languages).
3 / 16
🎯 Objectives & Research Questions
Primary Objectives
- Design a multi‑agent AI framework for research‑to‑software translation.
- Develop specialised agents for paper analysis, SRS, design, code, deployment.
- Implement a collaborative workflow with consistent high‑quality outputs.
- Evaluate effectiveness: completeness, consistency, accuracy, time savings.
Research Questions
- RQ1: How to design a multi‑agent framework for research‑to‑software?
- RQ2: What is the effectiveness in accuracy, completeness, and time savings?
- RQ3: How does it perform across ML, NLP, CV, SE domains?
- RQ4: What is the role of HITL in improving quality?
- RQ5: How to integrate Indian language support?
4 / 16
🏗️ System Architecture
graph TD
UI[User Interface Layer] --> Orchestrator[Agent Orchestration]
Orchestrator --> Agents[Agent Execution Layer]
Agents --> Providers[Provider Integration Layer]
Providers --> Tools[Tools & Services]
subgraph Agents
A1[Paper Analyzer] --> A2[SRS Generator]
A2 --> A3[Module Designer]
A3 --> A4[User Flow Designer]
A4 --> A5[SDLC Planner]
A5 --> A6[Code Generator]
A6 --> A7[Debugger]
A7 --> A8[Deployment Planner]
A8 --> A9[Final Reviewer]
end
subgraph Providers
P1[OpenRouter] & P2[NVIDIA Build] & P3[Claude] & P4[Ollama] & P5[KoboldCPP] & P6[Bhashini AI] & P7[Dev]
end
subgraph Tools
T1[OCR] & T2[Web Fetcher] & T3[Agent Runner] & T4[Mermaid Generator]
end
9 Specialised Agents – each responsible for one stage of the pipeline. Hierarchical coordination with forward/feedback flows.
5 / 16
🤖 Multi‑Agent Roles
| Stage | Agent | Output |
|---|
| 1 | Paper Analyzer | Structured Requirements |
| 2 | SRS Generator | IEEE 830 SRS |
| 3 | Module Designer | Module Design Artifacts |
| 4 | User Flow Designer | Sequence Diagrams |
| 5 | SDLC Planner | Sprint & Milestone Plan |
| 6 | Code Generator | Production Code |
| 7 | Debugger | Debugged Code |
| 8 | Deployment Planner | CI/CD + Infra Plan |
| 9 | Final Reviewer | QA Report |
Communication: Hierarchical coordination with forward/feedback flows.
6 / 16
🔄 Agent Collaboration
sequenceDiagram
participant PA as Paper Analyzer
participant SRS as SRS Generator
participant MD as Module Designer
participant UF as User Flow Designer
participant SDLC as SDLC Planner
participant CG as Code Generator
participant DB as Debugger
participant DP as Deployment Planner
participant FR as Final Reviewer
PA->>SRS: Structured Requirements
SRS->>MD: SRS Document
SRS->>UF: SRS Document
SRS->>SDLC: SRS Document
MD->>UF: Module Design
MD->>CG: Module Design
UF->>CG: User Flow
SDLC->>CG: SDLC Plan
CG->>DB: Code Files
CG->>DP: Code Files
DB->>FR: Debugged Code
DP->>FR: Deployment Plan
FR->>FR: QA Validation
Sequential flow with feedback loops for iterative refinement.
7 / 16
📐 Agent Class Hierarchy
classDiagram
class BaseAgent {
-config: Dict
-llm: LLMBackend
-logger: Logger
-prompt_template: str
-metrics: Dict
+process(input) Dict
+_prepare_prompt() str
+_parse_response() Dict
+_extract_context() Dict
}
class PaperAnalyzerAgent {
+process()
+_extract_text()
+_extract_algorithms()
}
class SRSGeneratorAgent {
+process()
+_load_template()
+_format_srs()
}
class ModuleDesignerAgent {
+process()
+decompose()
+define_interfaces()
}
class CodeGeneratorAgent {
+process()
+scaffold()
+implement()
}
class DebuggerAgent {
+process()
+scan_bugs()
+security_check()
}
BaseAgent <|-- PaperAnalyzerAgent
BaseAgent <|-- SRSGeneratorAgent
BaseAgent <|-- ModuleDesignerAgent
BaseAgent <|-- CodeGeneratorAgent
BaseAgent <|-- DebuggerAgent
All agents inherit from the abstract BaseAgent class.
8 / 16
🔌 Supported LLM Providers
| Provider | Free Tier | Indian Lang |
|---|
| OpenRouter | ✅ | ❌ |
| NVIDIA Build | ✅ | ❌ |
| Anthropic Claude | ❌ | ❌ |
| Ollama (local) | ✅ | ❌ |
| KoboldCPP (local) | ✅ | ❌ |
| Bhashini AI | ✅ | 24 languages |
| Dev (mock) | ✅ | ❌ |
🌐 Bhashini AI Integration – Translation, transliteration, TTS, ASR for Hindi, Tamil, Telugu, Bengali, Marathi, Gujarati, Kannada, Malayalam, Odia, Punjabi, Urdu, Sanskrit, and 12+ more.
USP: Pay only for what you use – free local providers (Ollama/KoboldCPP) enable fully offline execution.
9 / 16
🧑💻 Human‑in‑the‑Loop
| HITL Stage | Quality Improvement |
|---|
| Pre‑Inference Config | 8.2% |
| Post‑SRS Review | 12.4% |
| Post‑Code Review | 10.1% |
| Post‑Final Review | 7.6% |
⚡ USP Highlight · Data Attribution – every requirement, code block, and design artifact is traced back to the specific section, equation, or figure in the source research paper.
Local Deployment Ready: CLI + Docker + offline mode (Ollama/KoboldCPP). No mandatory cloud dependency.
10 / 16
📊 Methodology
Design Science Research – 6 phases: Problem ID → Objectives → Design → Demonstration → Evaluation → Communication.
Dataset
40 research papers (10 each: ML, NLP, CV, SE) from arXiv, IEEE, ACL, CVPR, ICSE.
Evaluation Metrics
- Completeness · Consistency · Accuracy · Executability
- Time Savings · Quality Rating (1‑5 Likert)
Statistical Validation: Paired t‑test, Cohen's d, ANOVA, Cronbach's α.
11 / 16
📈 Performance Results
| Metric | Collabuild | Manual | Improvement |
|---|
| Processing Time | 4.2 min | 180 min | 97.6% ↓ |
| SRS Accuracy | 78.4% | 92.1% | −13.7% |
| Code Compilation | 65.7% | 92.3% | −26.6% |
| Quality Rating | 3.9/5 | — | — |
Domain‑wise Accuracy
- ML: 82.1% · NLP: 79.4% · CV: 76.8% · SE: 75.3%
Cost: $0.50–$2.00 per paper (LLM API) – free with local providers.
12 / 16
⚖️ Comparison
| Framework | Research Input | SRS Gen | Multi‑Provider | Indian Lang |
|---|
| ChatDev | ❌ | Limited | ❌ | ❌ |
| MetaGPT | ❌ | Partial | ❌ | ❌ |
| iCoder | ❌ | Limited | ❌ | ❌ |
| Collabuild | ✅ | ✅ | ✅ (8) | ✅ (24) |
Why Collabuild stands out:
✅ Accepts research papers directly · ✅ Generates IEEE 830 SRS · ✅ 8 LLM providers · ✅ Bhashini AI (24 Indian languages) · ✅ Local deployment (Ollama/KoboldCPP) · ✅ Data attribution · ✅ Open‑source MIT License
13 / 16
👨🏫 Faculty Profiles
Dr. K. Udaya Sri (Guide)
Assistant Professor, Dept. of Computer Science, PB Siddhartha College
Ph.D. (Astrology) · M.Tech (CSE) · M.Sc (IT) · MBA (Finance)
15+ years teaching · 8 awards · 20+ international journals · SCOPUS ID: 57219411617
N.S. Sai Karun (Investigator)
Assistant Professor, Dept. of Data Science & AI, PB Siddhartha College
Expertise: Multi-Agent Systems, LLMs, Software Engineering
K. Gayathri (Investigator)
Assistant Professor, Dept. of Data Science & AI, PB Siddhartha College
Expertise: NLP, Machine Learning, AI-Augmented Development
K. Kishore (Investigator)
Assistant Professor, Dept. of Data Science & AI, PB Siddhartha College
Expertise: Software Engineering, Cloud Computing, DevOps
14 / 16
🔮 Future Work
Short‑term (0‑6 months)
- Multi‑modal processing (vision‑language for figures/diagrams).
- Fine‑tuning on academic paper datasets.
- Enhanced code generation with review agents.
Medium‑term (6‑12 months)
- Domain‑specific agents and validation.
- Cross‑lingual support for non‑English papers.
- Performance optimisation (parallel processing, caching).
Long‑term (12+ months)
- Active learning · Cross‑domain transfer · Production integration.
- Community development & industry collaboration.
15 / 16
📚 References
He, J. et al. (2024). LLM‑Based Multi‑Agent Systems for SE. arXiv.
Qian, C. et al. (2023). ChatDev. arXiv.
Hong, S. et al. (2023). MetaGPT. arXiv.
Wu, Q. et al. (2023). AutoGen. arXiv.
You, Z. et al. (2026). iCoder. ScienceDirect.
Daull, T. et al. (2025). CollabToolBuilder. WI‑IAT.
IEEE Std 830‑1998.
Copyright © 2026 Sai Karun Nandipati
MIT License · Open‑source · github.com/karun99/Collabuild
Department of Data Science & AI · PB Siddhartha College · Vijayawada · 2026
16 / 16