Company: Terasynth 

Solicitation Number: A244-065 

Principal Investigator: Ali Mahvan 

Business Official: Ali Mahvan, CEO

Submission Date: October 1, 2024

Volume 2: Technical Volume

1. Feasibility Documentation

Terasynth has successfully completed equivalent work to a Phase I effort, demonstrating the technical merit and feasibility of the proposed research. Our previous projects, such as the ReArmor AI-driven VR simulation application for PTSD exposure therapy and the AI-driven workflow optimization for autonomous Test Procedure Utility (TPU) document generation, and later guidance lent to Naval Air Warfare Training Systems Division (NAWCTSD), have laid a strong foundation for this endeavor. These projects involved developing AI models for decision support, analyzing vast quantities of technical manuals and IETMs, and working with complex, multi-domain data showcasing our capabilities in natural language processing, machine learning, data visualization, and AI-driven workflow optimization.

Terasynth has already conducted preliminary research and development specific to this proposal, including developing two prototype LLM applications: one for enhancing situational awareness by ingesting and analyzing OS and PAI data, and another for rapidly generating and evaluating potential COAs using an LLM-based algorithm. This preliminary work, while promising, allowed us to identify potential challenges in implementing LLM-driven COAA, such as mitigating bias in LLM-generated COAs, ensuring explainability and trust in LLM decisions, and enabling real-time adaptability in dynamic environments.

To address these challenges, we explored various solutions. These included using adversarial training and debiasing algorithms during the LLM's training process to minimize bias, integrating explainability methods like attention mechanisms to provide human-understandable rationales for the generated COAs, and developing a system architecture that allows for continuous learning and adaptation using reinforcement learning and human feedback.

Limited Trials and Prototype Capabilities

During the limited trials, Terasynth's prototype OSINT system exhibited a remarkable capacity for identifying subtle, nefarious cues embedded within the analyzed data. A deliberately curated subset of the sample data (1.3%) was seeded with indicators designed to trigger a "Red Flag Event," necessitating human intervention. 

In a particularly noteworthy instance, the system successfully flagged a user masquerading as an official government account, accurately identifying their content as a "Red Flag Event," while simultaneously recognizing the low source reliability due to the account's anonymous nature or lack of official verification (figure 1-a).

The AI system's ability to successfully flag each of these instances on its initial pass highlighted its potential for proactive threat detection and mitigation, even in the face of cleverly disguised or obscured malicious intent. This early success validates the underlying technical approach and sets the stage for further development and refinement.

Data Analysis and Output Capabilities

The prototype system, though in its nascent stage, demonstrated a proficiency in processing and analyzing OSINT data, extracting valuable insights from a complex and often noisy information landscape. The output data elucidates several key capabilities:

• Entity Recognition and Assessment: Leveraging advanced Natural Language Processing (NLP) techniques, the system can identify and classify entities within the data, ranging from individuals and organizations to locations and events. It goes beyond mere identification, assessing the relevance of each entity to a given scenario, gauging sentiment associated with the entity (positive, negative, or neutral), and quantifying the intensity of that sentiment. This capability allows for a nuanced understanding of the actors and their roles within the operational environment.

• Source Reliability and Information Accuracy: Recognizing the inherent variability in the quality of OSINT data, the prototype incorporates mechanisms to evaluate the reliability of information sources. This evaluation, combined with the system's own confidence assessment in the accuracy of extracted information, equips users with the critical ability to discern the trustworthiness of the generated insights.

• Statement Analysis: Going beyond simple information extraction, the prototype demonstrates the ability to analyze user statements, providing a concise summary of their intent or purpose. Crucially, it can also flag potentially concerning statements that may warrant further scrutiny by human analysts or authorities, acting as an early warning system for potential threats or risks.

• Human-in-the-Loop: Acknowledging the inherent limitations of even the most advanced AI systems, the prototype incorporates a crucial human-in-the-loop mechanism to ensure responsible and effective decision-making. When the system's confidence in its analysis falls below a predefined threshold, it proactively triggers a "humanLoopRequired" flag. This flag serves as a critical safeguard, preventing critical decisions from being made solely on the basis of automated analysis.

Furthermore, this human-in-the-loop mechanism acts as a dynamic gateway for integrating additional actions into the AI system's workflow. Upon triggering the flag, the system can initiate a range of responses, from simple notifications like sending an email alert to more complex actions such as launching an autonomous drone equipped with specific sensors to gather additional data at a designated location. This flexibility allows for seamless integration of human expertise and judgment into the AI-driven decision-making process, ensuring optimal outcomes even in situations where the AI's confidence is limited.

OSINT Entity Red Flagged for Content but Source Reliability Questioned

(figure 1-a)

AI Agent Nuance Detection

The preliminary analysis of the AI prototype's performance reveals its ability to identify and assess entities with a degree of nuance, even in its early stages of development. Specifically, entities classified as organizations exhibited a median relevance score of 10, indicating their high significance within the given scenario. The sentiment intensity associated with these organizations varied across a range of 9 and 10, suggesting a predominantly strong emotional tone, with the specific polarity (positive or negative) provided in the data for each instance of OSINT data tracked. Notably, the source reliability scores for these organizations were consistently high, ranging from 8 to 10, marking the system’s adeptness at identifying the credibility of the information sources used in the analysis. This combination of high relevance scores, varied sentiment intensity, and high source reliability demonstrates the systems capacity to not only identify key players but also to capture the subtle nuances in their roles and the sentiments surrounding them.

Median Entity Values for Entities Identified as Organizations

(figure 1-b)

By conducting these experiments, Terasynth further validated the feasibility of the proposed research, refine its technical approach, and build confidence in the potential of LLM-driven COAA to revolutionize decision-making in multi-domain operations.

2. Technical Proposal

Problem Statement:

In the increasingly complex and volatile battlespace of multi-domain operations, military decision-makers face mounting challenges in maintaining the initiative. Adversaries are adopting disruptive tactics to gain an edge, rendering traditional planning and execution methods less effective. The ability to rapidly identify and exploit fleeting opportunities, while mitigating risks, is crucial to achieving mission success.

Objectives:

This research aims to empower warfighters to proactively shape the battlefield by:

• Disruptive COA Generation: Leveraging Large Language Models (LLMs) to analyze vast amounts of OSINT and PAI, uncovering unconventional and unexpected COAs.

• OODA Loop Acceleration: Enabling rapid decision-making by automating key aspects of the Observe, Orient, Decide, and Act cycle.

• Strategic Surprise: Influencing both the rate and rhythm of engagement, creating opportunities to seize the initiative.

Expected Outcomes and Future Refinements

Terasynth's initial prototype demonstrated promising capabilities, it represents merely the first step in a journey towards a fully realized, operationally effective AI-driven decision support system with the potential of leveraging LLMs for situational awareness by extracting key information from OSINT and PAI data. Building upon this foundation, the proposed research will significantly expand the depth and breadth of data analysis, incorporating a wider array of data points and advanced extraction techniques. This enhanced capability will enable a more nuanced and comprehensive understanding of the operational environment. Future enhancements will focus on several key areas:

Database Schema Enhancement

The existing database schema will be augmented to include a richer set of data points, encompassing:

• Entity-Related Information: We will expand entity classification beyond basic types to include subtypes (e.g., military unit, political leader, NGO), affiliations, relationships between entities, and historical context.

• Event-Related Information: Event data will be enriched with additional attributes such as severity, casualty estimates, and potential triggers or consequences. We will also explore the integration of event forecasting models to anticipate potential future developments.

• Sentiment & Opinion: Sentiment analysis will be refined to capture nuanced emotions and opinions, going beyond simple polarity. 

• Geospatial Information: Geospatial data will be expanded to include terrain features, infrastructure, and points of interest. We will also leverage spatial reasoning capabilities to identify patterns and trends in the data.

• Temporal Information: Time-series analysis will be employed to track the evolution of events, entities, and sentiments over time, enabling the identification of emerging threats and opportunities.

Advanced Extraction Techniques

• Multilingual NLP: The system will be enhanced to process and analyze information in multiple languages, expanding the scope of data sources and enabling a more comprehensive understanding of the global information environment.

• Deep Learning for Information Extraction: We will leverage deep learning techniques, such as transformer-based models and graph neural networks, to extract complex relationships and patterns from unstructured data, enhancing the system's ability to identify hidden connections and generate actionable insights.

• Explainable AI (XAI): XAI techniques will be incorporated to provide transparency into the LLM's reasoning process, enabling users to understand the basis for the generated insights and fostering trust in the system's outputs.

Iterative Refinement and Validation

Throughout the development process, we will conduct extensive testing and evaluation, incorporating feedback from military subject matter experts to ensure the system's outputs are accurate, relevant, and actionable. We will also employ a human-in-the-loop approach, where human analysts can review and validate the LLM's findings, further enhancing the system's reliability and trustworthiness.

Technical Approach:

Terasynth's proposed prototype solution centers on the seamless integration of cutting-edge Large Language Models (LLMs) with our proven expertise in AI-driven decision support and data analysis. This synergy, honed through successful projects like ReArmor and our collaborations with industry and government organizations, directly addresses the solicitation's objectives. Key components and their alignment with past successes include:

• Data Ingestion and Enrichment: Our proprietary methods for filtering, refining, and looping data to refine AI analysis, demonstrated in prior work, will be leveraged to ensure the ingestion and enrichment of diverse OSINT and PAI sources is both efficient and effective. This will enable the system to generate a comprehensive, multi-dimensional operational picture, vital for identifying disruptive COAs.

• LLM-Driven COA Generation: Terasynth's experience in developing AI models for decision support and analyzing vast quantities of technical data, exemplified by the ReArmor project, directly translates to the employment of advanced language models for COA generation. Our fine-tuned LLMs, coupled with our understanding of military strategy and tactics, will identify unconventional COAs, assess feasibility, and provide invaluable insights to decision-makers.

• OODA Loop Integration: The automation of key OODA Loop steps will draw upon our proven capabilities in AI-driven workflow optimization, demonstrated in our work shared with NAWCTSD. By integrating real-time situation assessment, rapid COA evaluation, and proactive decision support, the system will empower warfighters to make informed decisions at the speed of relevance.

• User Interface and Visualization: A user-centric, intuitive interface will be developed, enabling warfighters to seamlessly interact with the system. The interface will provide interactive visualizations of COAs, allowing users to explore potential outcomes and trade-offs in a clear and concise manner. The system will also support natural language interaction, empowering warfighters to query the system and receive relevant information in real-time.

Methodology:

Our approach is grounded in an iterative and agile development process, incorporating continuous feedback and refinement. This methodology, honed through numerous successful projects, ensures adaptability and responsiveness to the evolving needs of the warfighter and the dynamic nature of the battlespace.

• Requirements Gathering and Analysis: We will collaborate closely with military subject matter experts, leveraging our experience in understanding and translating complex operational needs into actionable technical requirements.

• System Design and Development: Building upon the gathered requirements and informed by our past successes, we will design and develop the prototype solution. This phase will emphasize seamless integration of LLM capabilities with our proprietary data processing and analysis tools, ensuring interoperability with existing military systems.

• Testing and Evaluation: Rigorous testing in simulated and, where feasible, real-world scenarios will be conducted. User feedback will be actively incorporated to iteratively enhance system performance, usability, and effectiveness.

• Deployment and Transition: We will ensure a smooth transition to the operational environment by providing comprehensive training and ongoing support to end-users. Our commitment to collaboration and knowledge transfer will empower warfighters to fully leverage the system's capabilities.

Terasynth's technical approach and methodology are built upon a foundation of proven expertise and past successes. By leveraging our existing capabilities and incorporating user feedback throughout the development process, we are confident in our ability to deliver a prototype solution that exceeds expectations and empowers warfighters to proactively shape the battlefield.

Statement of Work

1. Introduction

This Statement of Work outlines the tasks, timelines, and deliverables for the development and enhancement of an AI-driven decision support prototype, aimed at empowering warfighters to proactively shape the battlefield in multi-domain operations. Building upon Terasynth's successful preliminary research and prototype development, this project will focus on refining and expanding the system's capabilities to deliver a robust, operationally effective solution.

2. Project Scope

The project encompasses the following key areas:

• Enhancement of Data Ingestion and Analysis: Expand the system's capacity to process and analyze diverse OSINT and PAI data, incorporating a wider array of data points and advanced extraction techniques.

• Refinement of Analysis Algorithms: Improve the accuracy, granularity, and robustness of entity recognition, sentiment analysis, and information reliability assessment.

• Integration of Explainable AI: Enhance the system's ability to provide clear and comprehensive explanations for its assessments and recommendations.

• Implementation of Real-Time Adaptability: Equip the system with continuous learning capabilities to adapt to evolving threats and changing circumstances.

• Rigorous Testing and Evaluation: Conduct extensive testing in simulated and real-world scenarios to validate system performance and ensure operational effectiveness.

3. Project Tasks

• Task 1: Advanced Data Ingestion and Enrichment (Months 1-3)

  • Expand the data ingestion framework to incorporate social network analysis, behavioral patterns, and predictive indicators.

  • Implement multi-lingual NLP and deep learning techniques for enhanced information extraction from unstructured data.

  • Develop sophisticated models for assessing source reliability and information credibility.

  • Augment the database schema to include richer entity, event, sentiment, geospatial, and temporal information.

• Task 2: Algorithm Refinement and Explainability (Months 4-6)

  • Refine analysis algorithms using cutting-edge machine learning techniques to improve accuracy and granularity.

  • Integrate explainable AI (XAI) methods to provide transparency into the LLM's reasoning process.

  • Evaluate the clarity and usefulness of generated explanations through user feedback and expert review.

• Task 3: Real-Time Adaptability and Continuous Learning (Months 7-9)

  • Implement continuous learning capabilities using reinforcement learning and human feedback.

  • Develop a system architecture that supports real-time adaptation to dynamic environments and evolving threats.

  • Evaluate the system's ability to incorporate new information and adjust its analysis in real-time.

• Task 4: Testing and Evaluation (Months 10-12)

  • Conduct rigorous testing in simulated and real-world scenarios, incorporating feedback from military subject matter experts.

  • Employ a human-in-the-loop approach to validate the LLM's findings and ensure alignment with operational needs and ethical considerations.

  • Refine the system based on testing results and user feedback.

• Task 5: Deployment and Transition (Months 13-15)

  • Prepare the system for operational use, ensuring compatibility with existing military infrastructure.

  • Develop comprehensive training materials and conduct training sessions for end-users.

  • Provide ongoing support and maintenance to ensure system functionality and address any issues.

4. Project Timeline

The project is estimated to take 15 months to complete.

5. Project Deliverables

• Enhanced AI-driven decision support system prototype

• Comprehensive documentation of system architecture, algorithms, and user interface

• Test plans, reports, and evaluation results

• Training materials and user documentation

6. Acceptance Criteria

The system will be considered acceptable upon successful completion of all tasks and deliverables, meeting the defined requirements, and demonstrating satisfactory performance in testing and evaluation.

7. Project Management

Terasynth will assign a dedicated project manager to oversee the project, ensuring timely completion of tasks, effective communication, and adherence to quality standards.

8. Intellectual Property

All intellectual property developed during the project will be owned by Terasynth Inc, with unlimited data rights granted to the government for use, safeguarding their investment and ensuring their rights to the resulting innovation.

Team Qualifications:

Terasynth has assembled a multidisciplinary team with extensive experience in AI/ML, defense applications, and software development. Key personnel include:

• Ali Mahvan, US Citizen, Project Manager: Ali's early work with AI, predating the ChatGPT craze, has positioned him at the forefront of data-driven workflow innovation. He pioneered integrations like real time internet lookup (utilizing SerpAPI Google Search), limited AI agent memory retention, real-time data, and ease of accessibility - features now standard in top-tier consumer AI software. This demonstrates his forward-thinking approach and significant impact on making AI technology more accessible and practical for everyday use.

• William Taubenheim, US Citizen, Technical Lead: Will Taubenheim played a pivotal role in the development of ReArmor, an AI-driven VR simulation application. This AI-driven VR simulation application, created under a sole-source contract with the University of Central Florida, utilized advanced natural language processing (NLP) algorithms to extract key data points from a clinical psychologist's narrative of a PTSD patient's trauma. These extracted data points were then autonomously translated into a 1:1 scale 3D virtual environment within Unreal Engine, requiring virtually no user interaction in the construction process. Will Taubenheim's expertise was crucial in implementing this sophisticated pipeline, ultimately empowering clinicians to leverage VR for exposure therapy in PTSD treatment.

• William Snook, US Citizen, Project Subject Matter Expert, former Director of Plans and Policy, Joint Interagency Task Force South and Director Commander's Planning Group, US Army: William Snook's extensive experience in senior-level military planning and strategic decision-making positions him as a valuable subject matter expert for this solicitation. His decade-long career in training management and strategic planning, coupled with his most recent role as Director of Plans and Policy for Joint Interagency Task Force South, demonstrates his deep understanding of the complexities and challenges of multi-domain operations. His proven ability to collaborate with decision-makers, develop and implement strategic plans, and optimize resource allocation directly aligns with the objectives of this project, which aims to empower warfighters with AI-driven decision support for rapid and accurate decision-making in dynamic battlefield environments. Furthermore, his experience in navigating nuanced scenarios and managing risk effectively will be instrumental in ensuring the proposed solution meets the stringent requirements of military operations.

• Company: Terasynth has established itself as a leader in the application of AI workflows within industry organizations serving government sectors. Notably, we have made significant strides in the following areas:

• Simulation Training Enhancement: We've developed AI-driven workflows to delineate step-by-step procedures for simulation training, including Test Procedure Utility documents. These documents are generated through AI consumption and vector embedding of extensive technical manuals and IETMs, enabling the AI agent to navigate maintenance processes and access a tool bank with relevant tools for each stage of work.

• Collaboration with Naval Air Warfare Training Systems Division (NAWCTSD): We've actively engaged with NAWCTSD, guiding their team through the technical complexities of integrating Large Language Models (LLMs) into dynamic workflow procedures. This collaboration aimed to expedite the RFP development process for an Other Transactional Authority (OTA) contract's technical requirements.

Our team's collective expertise positions us to successfully execute this project and deliver a solution that meets the Army's needs.

Volume 5: Supporting Documents

The following supporting documents are attached as additional documents to the solicitation.

• Commercialization Plan:
  A244-065 SBIR 24.4 Large Language Model (LLM) Course of Action Analysis (COAA) .pptx

• Statement of Work (SOW):
  A244-065 SBIR 24.4 SOW