We write history — and we are part of it.

February demonstrated PrimeForge's evolution toward Agent-Native capabilities with the launch of **PrimeForge / PonyBunny** — an open-source project aligning PrimeForge's platform with the emerging AI agent paradigm. Triggered by the explosive popularity of ClawdBot (OpenClaw) in late January, PonyBunny was designed to provide controlled, transparent, auditable agent execution — building on PrimeForge's existing ILP routing and Proof Chain infrastructure. Meanwhile, **PrimeForge / Reputra** was nearing completion of its AI-native rebuild for the Chinese market (20+ subscribers, £20/month), and **PrimeForge / ProperSorted** was announced for March launch targeting UK landlords facing HMRC's MTD deadline. OpenCode was discontinued this month, and the Cursor Pro subscription was not renewed — **Claude Code** had proven significantly superior for the founder's workflow.

November celebrated the commercial launch of **PrimeForge** and the acquisition of its first enterprise customer — a high-volume logistics/recruitment firm. The platform's value proposition was validated by immediate, measurable results: reducing the client's API costs by 40% in Week 1 through the LP Router's intelligent provider selection. A "PrimeForge Developer" subscription tier was introduced, positioning the platform not just as an API but as a compilation target for AI-native applications. This was the beginning of the stable enterprise revenue stream the founder had been building toward since the January 2025 pivot.

September introduced the "Proof Chain" — a cryptographic audit mechanism that made compliance machine-verifiable. By implementing an append-only hash chain linking inputs, outputs, and timestamps for every execution step, the system provided a tamper-evident record of all routing decisions and outcomes. This ensured that the ILP router's dynamic nature did not compromise auditability — a critical requirement for compliance-focused applications. The Proof Chain became an additional patent claim, extending the invention's scope beyond routing optimisation into verifiable execution.

July was dedicated to formalising the routing mechanism for patent protection — translating the working code into a clean, defensible mathematical model. The invention was framed as a system that compiled GraphQL requests into a **Directed Acyclic Graph (Query Graph)** and computed an optimised execution plan using a real-time **Cost Matrix** derived from observability metrics. This formal specification aligned legal claims with actual system behaviour (IP-4: theoretical formalisation). This month also marked the adoption of **Claude Code Pro** (subscribed 13 July 2025), introducing CLI-based agentic development — a qualitative leap beyond Cursor Pro's IDE-centric approach.

May focused on validating the ILP router's reliability through systematic simulation testing. A simulation harness was built to generate thousands of synthetic requests with realistic tail latency, intermittent failures, and varying query depth — providing empirical evidence that the ILP planner could degrade gracefully under load. On the platform building line, the AI-native architecture continued maturing. This month also saw the adoption of **ChatGPT Plus** (subscribed 6 May 2025), adding another AI tool to the development workflow alongside Cursor Pro.

March addressed two parallel tracks. On the **platform building line**, the AI-native architecture continued taking shape with MCP-based tools being developed and tested. On the **IP research line**, the founder travelled to Shanghai for consultations with Professor Wang Dong (SJTU) and platform architects from large-scale payment/cloud systems — confirming that the routing problem should be solved with mathematical optimisation, not hand-coded rules. The result was a concrete blueprint: replace the rule engine with an ILP (Integer Linear Programming) based planner fed by runtime metrics. This completed IP-3 (algorithmic direction confirmed) and initiated IP-4 (core innovation implementation). This month also marked the adoption of **Cursor Pro** (annual subscription from March 2025) — a generational upgrade from Copilot, enabling codebase-aware AI assistance.

January 2025 was the defining month in DarkhorseOne's history — the **only technology route-level pivot** the company had ever made. On 20 January, DeepSeek released R1, a reasoning model with chain-of-thought capabilities comparable to OpenAI's o1 — free, open-weight, with native Chinese language support and transparent thinking. **Within one week, nearly the entire Chinese user base of DarkhorseOne HR migrated to DeepSeek.** The AI HR assistant business in China was effectively terminated overnight. In response, the founder made the decision to abandon the traditional SaaS + AI Bot model entirely and rebuild the entire system architecture as AI-native infrastructure, using Anthropic's newly released MCP (Model Context Protocol) as the foundational interface. This was the transition from Era One (DarkhorseOne HR) to Era Two (what would eventually become the PrimeForge platform).

November brought two significant developments. Operationally, the company relocated HQ to 27 Castle Street. Technically, the TypeScript-based query parser was rewritten in **Go** after performance profiling revealed serialisation overhead as a bottleneck — improving performance by roughly an order of magnitude. An experiment using OpenAI's **o1** model as a routing engine produced high-quality plans but was impractical due to multi-second inference latency. Most importantly, this month confirmed the **IP-3 phase** of the patent research line: the recognition that static routing rules were fundamentally insufficient, and the problem should be treated as a mathematical optimisation problem.

September witnessed a critical stability incident in DarkhorseOne HR's backend. Synchronous processing of mixed-latency data sources caused thread pool exhaustion and service unavailability when users retried during slow scraping operations. The root cause was the system's "wait for all" logic — a single slow scraper (8-15s) blocked fast API responses (200ms), triggering user refreshes that compounded into cascading failures. This incident was a turning point: it crystallised the need for dynamic query planning and exposed the fundamental limitation of static routing in a system with diverse data source characteristics (IP-1: the architecture pressure point that would drive the patent research line forward).

July expanded DarkhorseOne HR's compliance capabilities by integrating multiple UK government and private data sources — Companies House, The Gazette, ICO Register, and HM Land Registry — to build a Unified Corporate Compliance Graph. To manage the growing complexity of connecting disparate data sources (fast APIs, slow XML feeds, headless scrapers), a **GraphQL Federation** architecture was adopted as a unified query layer. While the connections were successfully established, a critical **latency mismatch** between fast APIs (200ms) and slow scrapers (8-15s) emerged as an architectural concern. This integration work directly exposed the engineering complexity that would seed the patent research line (IP-1/IP-2).

May 2024 brought the most significant technical disruption DarkhorseOne HR had faced since its inception. OpenAI's release of GPT-4o (13 May) — with native audio processing, image understanding, and dramatically improved reasoning — rendered months of engineering investment in custom RAG pipelines, OCR processing, and document extraction economically unviable. This was not a product pivot but a **forced technology stack replacement within the same product direction**: the DarkhorseOne HR product continued, but the underlying technical components had to be rebuilt around model-native capabilities. Across the industry, a wave of RAG-focused startups collapsed as their core differentiation was absorbed into the foundation model.

March initiated a new R&D track within DarkhorseOne HR: Automated Job Description (JD) Generation for UK SMEs. This was a natural feature extension of the product — helping clients not just manage HR compliance but also produce compliant hiring documentation. The primary technical effort involved developing a custom data scraper (migrated to TypeScript) that successfully harvested 2,500 anonymised job descriptions from Indeed, laying the foundation for AI-powered JD generation. Meanwhile, the IP research line quietly began: the growing number of external API integrations was starting to expose system complexity challenges (IP-1 — problem discovery).

January was a pivotal month for DarkhorseOne HR on two fronts. Commercially, the company secured its first significant client contract — a one-year customised development engagement built on the DarkhorseOne HR platform, validating both the market need and the platform's flexibility. Technically, extensive testing of open-source models (Llama 2, Mistral, Mixtral) as cost-effective alternatives to GPT-4 revealed critical reliability issues for HR use cases. The conclusion was clear: open-source models were not yet production-ready for this domain, confirming the decision to continue with GPT-4 + RAG as the primary technical path for DarkhorseOne HR.

November was dedicated to intensive engineering on DarkhorseOne HR. The strategic focus was "Schema Design First" — given the complexity of UK employment law, the data model had to be rigid enough for compliance but flexible enough for AI operations. This month also marked the adoption of **GitHub Copilot** (subscribed 5 November 2023, $10/month), which significantly accelerated the single-developer workflow by providing intelligent code completion and boilerplate generation across the TypeScript stack.

July 2023 marked the founding of **DarkhorseOne** — a company created to bring over a decade of Chinese HR industry expertise and 21+ years of IT development experience into the UK market, powered by the emerging wave of generative AI. The company was officially incorporated on **13 July 2023** as a Private Limited Company in England and Wales. At that point, the founder was still based in China, preparing for relocation to the UK. The company remained dormant during this period.