#50 - Signal Management / Jobwatch / Kono Knowledge Graphs

Link to Folder with Everything

2026.08.12.officehours.50

Recording of the Video

Watch on Google Drive

Summary (by Gemini)

Session focused on methodologies for signal management via deterministic crawling tools and knowledge graph-based agent systems.

Job search automation tools

Participants explored an open-source tool utilizing deterministic code to crawl job data and rank opportunities via AI agents. This system focuses on scalability and cost-efficiency by prioritizing standard coding over constant agent reliance.

Knowledge graph system implementation

Discussion shifted to a knowledge graph product designed to ground AI agents in verifiable facts and minimize hallucinations. This approach utilizes Resource Description Framework triplets to create a persistent, structured schema.

Workflow integration and evolution

The final segment covered the integration of data mining techniques with knowledge graphs to track market intelligence. Future development aims to combine these data streams for predictive pattern recognition and accessibility.

Next steps

• [Jonas] Send Job Digest: Email the generated job opportunities to Nadeen. Provide the final ranked list as discussed in the meeting.
• [Rahul] Present Code Fork: Demonstrate the custom modifications made to the jobwatch repository. Explain the specific additions to the original project during a future meeting.
• [Rahul Singh] Share Readme Files: Share the readme information for all repositories utilized in the deck presentation with the group.
• [Rahul Singh] Perform QA: Conduct quality assurance on the collected market watch data to improve content curation accuracy.
• [Rahul Singh, Obioma Nwabueze Anomnachi] Connect Data: Integrate the generated data from external sources into the Kono knowledge graph for improved visualization and prediction.

Details

• Introduction and Remote Work Evolution: The participants discussed the evolution of remote collaboration tools, reflecting on the historical use of software like Skype and how current AI capabilities allow for efficient knowledge work across geographic distances.
• Office Hours 50 Agenda: Rahul Singh introduced the fiftieth session of "Office Hours," which focused on two distinct methodologies for managing information signals and discovering opportunities: Jonas von der Heyden's "Jobwatch" open-source tool and Obioma Nwabueze Anomnachi's knowledge graph-based signal management system (00:16:51).
• Jonas von der Heyden's Background: Jonas, a recent computer science PhD graduate in cryptography, described developing the jobwatch tool after finding traditional job platforms inefficient and overly dependent on advertising models (00:21:58). To address this, Jonas created an AI-powered agent to scrape job data, eventually transitioning to deterministic code for crawling to improve cost-efficiency and avoid AI usage limits (00:23:09).
• Obioma Nwabueze Anomnachi's Background: Obie, a data engineer with a history in machine learning, explained their transition from creating text classifiers to focusing on workflow automation and AI agents. Obie noted a preference for practical task execution and utility over running agents constantly (00:26:31).
• jobwatch Live Demonstration: Jonas demonstrated the jobwatch tool using a curriculum vitae provided by Nadeen Siddiqui to identify employment opportunities (00:30:06). The tool extracted preferences—including "AI product leadership" and "remote-friendly US roles"—from the document and initiated a search process, identifying relevant companies such as Anthropic and OpenAI (00:32:37).
• jobwatch Architecture and Workflow: Jonas explained that the jobwatch architecture employs deterministic code for website crawling to minimize costs, while specialized AI agents in an integration queue handle website parsing code (00:40:28). The system is designed to create a flywheel effect, where user contributions of integration code improve the tool for the entire community (00:43:20). A secondary AI ranker scores job listings against user preferences to provide a curated digest via email or messaging (00:44:28).
• Scaling and Future Potential of jobwatch: Jonas detailed the strategy of prioritizing deterministic code over AI agents for crawling to ensure scalability, lower costs, and reduce the likelihood of being blocked by target websites (00:47:48). Future developments may include a hosted version for non-technical users and applications for tracking other data, such as research grant applications or the price of electronic items (00:49:12).
• Q&A - Job Matching Quality: Lea Bauer questioned the ability of the tool to distinguish between a company's public branding and the reality of a role (00:51:42). Jonas clarified that the tool utilizes a two-step process: an initial relevance filter and a ranking agent that evaluates job listings against specific criteria defined in a "preferences.md" file (00:54:00). Jonas also confirmed that the tool operates automatically using a cron job, running daily on the user's computer to provide updates without repeating previously seen opportunities (00:56:24).
• Signal Flow and Knowledge Graph Approach: Obie introduced an alternative methodology called "Signal Flow," which emphasizes an agent-reliant approach rather than purely deterministic code (01:00:18). This system processes information through stages: explore, seek, gather, ingest, organize, understand, and generate, utilizing a knowledge graph to maintain structured context and data relationships (01:01:33).
• Libra Chat and Tool Integration: Obie demonstrated the use of "Libra Chat," an open-source interface layer, to manage agents, skills, and Model Context Protocol (MCP) servers (01:04:21). Rahul Singh highlighted Libra Chat as a solution for "AI sovereignty," enabling users to maintain off-grid or air-gapped AI environments (01:05:34).
• Knowledge Graph Fundamentals: Obioma Nwabueze Anomnachi explained the distinction between standard databases, graph databases, and knowledge graphs (01:12:49). While a standard database stores key-value pairs and a graph database maintains nodes and edges, a knowledge graph utilizes RDF (Resource Description Framework) triplets—groups of one edge and two nodes—to provide an English-readable structure of connections that allows for logical deduction (01:14:43).
• Introduction to Kono: Obioma Nwabueze Anomnachi introduced Kono, a knowledge graph product designed for AI agents to mitigate hallucinations—intrinsic (content factually incorrect relative to the source) and extrinsic (a fabricated source) (01:17:48). Kono addresses these issues by providing provenance for every claim, linking it to a specific document and page for validation (01:19:06).
• Knowledge Graph vs. Graph RAG: While Graph RAG retrieves text snippets to answer specific questions without inherent structure, Kono establishes a persistent schema that grounds connected agents in shared logic (01:24:12). Unlike Graph RAG, which typically functions as a temporary tool for answering a single user request, Kono is designed as a living entity where new services connect and update data over time (01:25:30).
• Kono Ingest and Memory Capabilities: Kono's ingestion process extracts information into RDF triplets, followed by de-duplication and a review queue for items that do not fit the established ontology (01:26:50). Kono also offers agent memory built on a knowledge graph of interrelated concepts rather than simple file storage, allowing agents to reconcile contradictory information by compacting memories into a coherent state (01:28:08).
• Iterating on Drive Code: Rahul Singh presented an iteration on Jonas von der Heyden's "Drive" code, utilizing a "signal flow" approach that organizes work into six stages: explore, seek, gather, organize, understand, and generate (01:33:38). By applying these verbs as a discipline, the code was mapped to a "Tech Observer" tool, which organizes and executes data mining tasks for tracks such as AI architecture and leadership (01:35:11).
• Methodology and Software Distribution: The systematic approach uses an "archetype" skill to reverse-engineer software into a layer-by-layer file stack, allowing for regeneration of the system, running on "TechMD," an open-source distribution for managing AI fleets (01:37:34). Rahul Singh noted that the development relied on Claude code rather than manual coding, and used the adapted code to track AI topics, market watch data, and job watch listings, publishing results to spaces accessible via Libra Chat or a knowledge graph (01:39:23).
• Collaboration and Future Outlook: The participants discussed the potential to integrate Rahul Singh's collected data into Obioma Nwabueze Anomnachi's Kono system to identify predictive patterns (01:43:09). The meeting concluded with a preview of upcoming office hours featuring speakers Michael Mayernick and Cherry Mountain (01:44:07).