EvidenceWeaver

Inspiration

EvidenceWeaver was inspired by both personal experience and a growing global problem.

I work full time in customer support for a telecommunications provider, where I regularly encounter customers who have become victims of fraud. While "pig butchering" scams initially sparked the idea for this project, the scams we see every day span a much broader spectrum—from stolen phones and SIM swap attempts to account takeovers, reshipping scams, compromised bank accounts, and sophisticated social engineering attacks. Regardless of the specific technique, the outcome is often the same: customers can lose thousands or even tens of thousands of dollars, along with countless hours trying to recover their accounts and finances.

One of the most difficult aspects of these situations is that, by the time customers contact us, the scam has often occurred outside of our systems. Although we can help secure their telecommunications account, the fraud itself frequently involves third-party messaging platforms, cryptocurrency exchanges, banking services, fake investment websites, or other external services that are beyond our ability to monitor or prevent. Too often, there is little we can do beyond helping limit any additional damage.

Watching these situations unfold highlighted how fragmented digital investigations have become. A single case may involve hundreds of chat messages, multiple online identities, cryptocurrency wallets, financial transactions, phone numbers, websites, and accounts spread across numerous platforms. Organizing all of that information into a coherent investigative picture is both time-consuming and difficult.

EvidenceWeaver was created to explore how AI can help solve that problem—not by replacing investigators or making accusations, but by helping organize evidence, reveal connections, and preserve the transparency needed for investigators to understand exactly how every conclusion was reached. Rather than asking people to simply trust AI, our goal is to make AI's reasoning traceable back to the evidence itself.

What it does

EvidenceWeaver is an AI-assisted investigation workspace designed to help organize complex digital evidence into a coherent investigative case. It is also the early foundation for a more advanced investigation platform planned for the future.

Users can upload or paste evidence such as chat conversations, documents, images and other information containing files, after which the system extracts key entities including people, aliases, cryptocurrency wallets, websites, phone numbers, financial transactions, and important events. The extracted information is organized into a chronological timeline and an interactive relationship graph, allowing investigators to quickly understand how different pieces of evidence connect.

Unlike traditional AI assistants, every extracted fact and analytical observation remains linked back to its original source material. Investigators can review, confirm, edit, or reject AI-generated findings before generating an evidence-backed case summary, helping maintain transparency and trust throughout the investigative process.

How we built it

EvidenceWeaver was developed through a collaborative workflow using ChatGPT and OpenAI Codex 5.6. ChatGPT was used during the planning phase to refine the project concept, investigative workflow, data model, and overall architecture before implementation began. Development then shifted primarily to Codex, which was used to iteratively build, test, and refine the application. In addition to generating code, Codex also served as a technical mentor throughout development, answering questions about any processes I was unfamiliar with. This allowed me to make informed design decisions while steadily improving my understanding of the technologies involved.

Technically, EvidenceWeaver was built as a modern web application using a React and TypeScript frontend with a lightweight backend connected to OpenAI's APIs for structured information extraction and analysis. While the AI pipeline has been integrated into the application's architecture, the hackathon submission uses three fictional investigation cases with synthetic data and a self-contained demonstration mode, allowing judges to explore the complete workflow without requiring live API access.

Rather than relying on a single prompt, EvidenceWeaver uses a staged AI pipeline consisting of evidence normalization, structured entity extraction, event extraction, relationship mapping, and evidence-backed case summarization. This modular approach improves consistency, explainability, and allows each stage to build upon the previous one while preserving traceability back to the original evidence.

Challenges we ran into

One of the biggest challenges was balancing AI automation with investigator trust. Large language models are excellent at finding patterns, but investigations require precision, traceability, and careful handling of uncertainty. We spent considerable effort designing prompts and data structures that distinguish between explicit evidence, inferred relationships, and higher-level analytical observations, and insured that this information about the level of confidence of various relationships was always directly relayed to the investigators.

Another significant challenge was representing real investigations as they actually occur. Evidence rarely arrives in a neat chronological order—it is scattered across chat conversations, financial records, cryptocurrency transactions, websites, and multiple online identities, often with conflicting or incomplete information, or may reference unidentified people and accounts that cannot immediately be linked together. We needed to organize this information without forcing premature assumptions. We designed EvidenceWeaver to preserve that uncertainty while progressively organizing the evidence into timelines, entity relationships, and investigation summaries that remain traceable back to their original sources, able to provide more and more clarity to a case as more evidence and information becomes available.

Finally, we deliberately designed EvidenceWeaver as a decision-support tool rather than a decision-maker. From the outset, we viewed AI as a tool for augmenting human investigation rather than replacing human judgment. EvidenceWeaver was intentionally designed to organize evidence, surface meaningful connections, and preserve transparency while leaving actual investigative conclusions to qualified human investigators. Every AI-generated insight is meant to assist human investigators, not replace them.

Accomplishments that we're proud of

One of our biggest accomplishments was demonstrating that AI can be used to build tools that make AI itself more transparent. Rather than functioning as another chatbot that produces answers without explanation, EvidenceWeaver was designed around evidence provenance, allowing investigators to trace every extracted entity, relationship, and analytical observation back to the original source material. That commitment to transparency became one of the defining principles of the project.

We're also proud of creating a platform that focuses on augmenting human investigators rather than replacing them. Every major design decision—from separating factual extraction from analytical interpretation to emphasizing human review—was made to ensure that AI supports investigative workflows without becoming the final decision-maker.

From a development perspective, we're proud that EvidenceWeaver demonstrates what is possible through AI-assisted software engineering. As someone without a traditional software development background, I was able to take a concept born from real-world customer experiences, collaborate with ChatGPT and Codex to refine the architecture and implementation, and rapidly build a sophisticated investigative prototype that would have otherwise been well beyond my technical abilities.

Finally, we're excited that the underlying framework extends well beyond pig-butchering scams. The same architecture could support investigations involving financial fraud, cybercrime, insurance fraud, account takeovers, identity theft, and many other forms of complex digital investigations.

What we learned

One of the biggest personal learning experiences during this project was gaining a much deeper understanding of how modern AI applications are actually built. Before starting EvidenceWeaver, concepts such as API integration, API keys, request pipelines, and deployment were largely abstract ideas. Working through those pieces with Codex transformed them into practical knowledge and gave me a much better understanding of how AI-powered applications communicate with external services.

Perhaps the biggest surprise, however, was not the technology itself—it was the development process. I expected to spend the majority of the hackathon working through countless small implementation details and debugging issues. Instead, by combining careful planning with iterative collaboration using ChatGPT and Codex, EvidenceWeaver evolved from an idea into a functional application far more quickly and smoothly than I had anticipated. Ironically, previous personal projects such as building a rich text editor required considerably more back-and-forth than developing this much larger investigation platform.

The experience fundamentally changed my perspective on AI-assisted software development. Rather than programming syntax being the primary obstacle, the real challenge became clearly defining the problem, designing intuitive workflows, and making thoughtful architectural decisions. Once those pieces were well understood, implementation became remarkably efficient. It reinforced the idea that AI is not simply accelerating coding—it is lowering the barrier between recognizing problems worth solving and building solutions that can make a real difference

What's next for EvidenceWeaver

EvidenceWeaver was designed as the foundation for an increasingly more capable and comprehensive investigation platform.

Future development will focus on expanding beyond single-case analysis into cross-case intelligence, allowing investigators to identify shared wallets, aliases, communication patterns, and scam infrastructure across multiple investigations.

We're also interested in adding behavioral pattern analysis that can recognize common scam playbooks and scripts while clearly communicating uncertainty and providing supporting evidence. Longer-term, we'd like to explore tools that help identify indicators of human trafficking and coercion within scam-center communications, providing investigators with additional context while remaining grounded in evidence.

Looking further ahead, we envision EvidenceWeaver becoming an optional investigation platform that can integrate with organizational communication systems. Rather than requiring investigators to manually collect and organize customer interactions, EvidenceWeaver could process customer service chats, call transcripts, emails, and other authorized communication records as they are generated, automatically extracting relevant entities, identifying recurring fraud indicators, and building evidence-linked case timelines. By continuously organizing information across many customer interactions, investigators could identify emerging fraud campaigns, connect related incidents earlier, and spend more time investigating meaningful leads instead of manually assembling evidence.

An even longer-term vision is to move beyond investigation and into prevention. As organizations gain confidence in evidence-backed detection, EvidenceWeaver could provide early-warning notifications when ongoing customer interactions exhibit patterns consistent with known fraud campaigns. Rather than making automated decisions or blocking legitimate activity, the system would surface evidence-backed risk indicators to employees or customers, encouraging additional verification before sensitive actions—such as account changes, device replacements, wire transfers, or cryptocurrency transactions—are completed. Helping stop fraud before a victim suffers financial loss has the potential to create an even greater impact than investigating cases after the damage has already occurred.

Ultimately, our goal is to build an investigation platform that helps analysts spend less time organizing information and more time understanding it—using AI to reveal connections while ensuring every conclusion remains transparent, reviewable, and backed by evidence.

Built With

• artificial-intelligence
• consumer-protection
• cryptocurrency
• cybersecurity
• data-visualization
• decision-support
• digital-forensics
• entity-extraction
• evidence-analysis
• explainable-ai
• financial-crime
• fraud-detection
• fraud-prevention
• generative-ai
• human-in-the-loop
• information-extraction
• knowledge-graph
• llm
• openai
• public-safety
• relationship-mapping
• scam-detection
• threat-intelligence
• timeline-analysis
• trust-and-safety