PakForge

DecisionForge is an agentic AI investigation engine built for enterprise decision-making. Unlike standard RAG or chatbots that answer immediately from surface-level retrieval, DecisionForge behaves like a human analyst — it plans a multi-step investigation, retrieves and reranks evidence across 65 cross-referenced enterprise documents (incident reports, ADRs, policies, Jira tickets, Slack threads, meeting notes), and reasons step-by-step until it has genuinely exhausted the evidence. The core loop: Plan → Retrieve → Rerank → Evaluate → Continue or Stop — repeating across 3–4 document hops until the Evaluator decides evidence is complete. Every number the system produces (confidence score, evidence found/used/rejected/missing) is computed deterministically in Python — never hallucinated by the model. The final output is a structured ACCEPTED / REJECTED / NEED_MORE_CONTEXT verdict with full citations, a reasoning trace, a generated Executive Memo, and a Follow-up Ticket. Built on Vultr Serverless Inference using two specialized models: Qwen/Qwen3.6-27B for planning, evaluation, and decision reasoning — and VultronRetrieverCore-Qwen3.5-4.5B for semantic evidence reranking. Results stream live to the frontend via SSE with a real-time investigation graph, evidence panel, and downloadable artifacts.