Terminus was built during a 48-hour hackathon as a working prototype of a trustworthy, locally deployable AI system. It combines Retrieval-Augmented Generation (RAG) with a local LLM (Ollama3) to ensure every response is backed by verifiable peer-reviewed research.
Unlike generic AI assistants, Terminus never fabricates information. Every answer links directly to its original source, complete with author, journal, and year.
Core Outcome:
100% citation accuracy
0% hallucination rate
Fully functional local deployment on standard laptop hardware
Background
This project began through the Rockwell Fellowship, where I was paired with Kris Rockwell as a real-world client. The fellowship aimed to bring together entrepreneurship, product development, and new technologies through hands-on work with real client problems.
During discovery meetings, Kris expressed a need for AI systems that could be trusted in secure, research-heavy environments, settings like healthcare, engineering, and academia where sending data to cloud APIs wasn’t possible.
He highlighted five critical challenges:
Privacy & Security: Sensitive data couldn’t be processed in the cloud.
Reliability: Most AI systems hallucinate or provide unverifiable claims.
Domain Specificity: General-purpose models lack research-level depth.
Cost Control: API-based models were too expensive to scale.
Offline Access: Research teams needed tools that run locally.
When the university announced the 48-hour hackathon challenge: “AI Reliability Through Peer-Reviewed Research Integration”, it was a perfect alignment. The challenge directly matched Kris’s real-world need.
Goal
To create a system that restores trust in AI by combining:
Local Privacy – Runs entirely offline via Ollama3.
Research Integrity – Sources only peer-reviewed publications.
Traceable Results – Every claim is verifiable through citation validation.
Transparency – Admits when information isn’t available, instead of guessing.
Accessibility – Easy to deploy for universities, research labs, and small teams.
Research
I conducted interviews, literature reviews, and user validation to confirm the need.
Findings included:
80% of AI users express concerns about misinformation and hallucinations (Forbes).
Researchers lose time verifying AI results manually.
Small research firms can’t afford enterprise AI licenses or external APIs.
Use Cases Identified
University labs conducting literature reviews.
Small R&D teams validating environmental materials.
Professors and students seeking reliable AI tutoring tools.
Solution Space
Terminus integrates modern AI architecture with research-grade verification.
Architecture Highlights
Document Ingestion Layer: Extracts and embeds peer-reviewed papers for semantic search.
RAG Engine: Retrieves the most relevant sections of text.
Ollama3 Core: Generates responses strictly from retrieved documents.
Citation Validator: Confirms every claim matches the cited paper.
User Interface: Displays clear answers with linked citations and confidence scores.
Tech Stack:
Ollama3 (Local LLM)
LangChain (RAG pipeline)
ChromaDB / FAISS (Vector database)
Python + Gradio Interface
Sentence Transformers for embeddings
Result:
A system that delivers academic-quality answers, complete with sources and zero fabrication.
Final Product & Learning
The final prototype was fully functional and demonstrated during the hackathon:
Query Example: “What is the environmental impact of PLA vs traditional plastics?”
Response: Terminus cited verified studies from Nature Materials and Journal of Applied Polymer Science, each cross-checked for accuracy.
Verification Outcome:
Citation Accuracy: 100%
Hallucination Rate: 0%
Relevance: 100%
Evidence Strength: 100% peer-reviewed