--- title: "Technical Roadmap" subtitle: "Kav AI Platform Milestones & Engineering Deliverables" --- This document outlines the technical development milestones, track-specific capabilities, and future engineering horizons for the Kav AI Platform (KAP). Aligned with Section 4 ("The Journey So Far") of the living Product Requirements Document (PRD), this roadmap details the transition from initial platform bootstrap to a mature, closed-loop, and compliance-ready industrial integrity solution. --- ## Completed Milestones ### Milestone M0: Platform Foundation (Jul 2025) * **Status:** Complete / Delivered * **Core Deliverables:** * Basic 3D visualization and rendering viewer (CesiumJS foundation). * Standard multi-tenant authentication and user access controls. * Image gallery and anomaly validation panel for operators. * **Validation:** Verified with a first real-world industrial dataset consisting solely of RGB imagery. ### Milestone M1: AI Foundation (Dec 2025) * **Status:** Complete / Delivered * **Core Deliverables:** * Multimodal AI analytical pipeline and unified NLP interface (atlas/argus/orion/kawa). * Automated agentic task coordination and scheduling. * Machine vision engine prototype capable of on-demand defect detection. * **Validation:** Verified with a richer, multimodal dataset containing thermal imagery and structured gas sensor readings (temperature, humidity, gas concentration). ### Milestone M2: App MVP (Mar 2026) * **Status:** Complete / Delivered (Released under `v0.2.0` on Apr 25, 2026) * **Core Deliverables:** * Unified single-screen operator interface combining the photorealistic 3D viewer and the AI Chat Copilot. * Extracted core component library and modular workspace architecture (DE/IE personas). * High-performance 3D Tile rendering, gas plume heatmap overlays, and standard input forms (KavAI Design System). * **Validation:** Validated against a dataset scanning 1,200 assets, identifying 37 thermal anomalies with 8 high-priority items representing $750K–$1.5M in avoided facility damage. --- ## Active Milestone ### Milestone M3: AI Q2 Delivery (Jun 2026) * **Status:** In Progress (Active shipment; target release is `v0.3.0` on May 27, 2026) * **Core Deliverables:** * **Contextual Data Chat Phase 0:** Conversational data querying supporting context-aware chips, workspace scoping, and multi-turn session state. * **Persona Workspaces:** Modular dual-route shell (Data Explorer & Integrity Engineer) with role-gated, persona-specific modules. * **Failure Recovery:** Resilient chat sessions — message queueing and retry across network interruptions and agent failures. * **Agent Evaluation Tests:** Automated end-to-end agent evaluation suite (correctness and regression) gating each release. --- ## Milestone M4: Persistent Sensing (Q3 2026) Q3 2026 deepens detection and engineering context across the industrial data ecosystem: * **Multi-Sensor Automated Detection:** AI-powered anomaly classification running continuously across OGI, calibrated thermal, and gas feeds, feeding directly into the core analytical chain. * **CAD (IFC) Ingestion & P&ID Tag-Linkage:** Open-standard engineering context — CAD geometry via IFC4 (BIM STEP) with logical P&IDs via DEXPI and deterministic dual-tagging. * **Geo-Tagged Spatial Anchoring:** Ingestion of precise GIS coordinates, automatically anchoring inspection campaigns and assets in the 3D space. * **Contextual Chat Phase 1:** Compound queries traversing multiple distinct data sources with human-in-the-loop (HITL) confirmation gates. * **Rescheduled from Q2 (M3 → M4):** Sensor-native agentic analysis, actionable insights, chat with 3D map, interactive overlays, and automated reports — deferred as M3 narrowed to contextual chat, persona workspaces, failure recovery, and agent evaluation. --- ## Milestone M5: Engineering Context & Enterprise (Q4 2026) The final quarter of 2026 is dedicated to enterprise-grade compliance, structural engineering ingestion, and deployment flexibility: * **OPC UA SCADA Connectors:** Read-only, real-time ingestion of time-series telemetry (vibration, pressure, temperature, and process historian logs), decoupling KAP from specific control-system vendors. * **3D CAD Model Overlay:** Overlaying actual engineering design models (IFC, RVT, or DGN formats) directly onto photorealistic Gaussian Splat models to perform as-built vs. as-designed deviation analysis. * **P&ID SQL Connector:** Direct read access to the plant's Piping & Instrumentation database to automatically populate the 3D model with fresh engineering metadata. * **SOC 2 Type II Certification:** Audit completion to meet the rigorous data governance requirements of major global operators. * **Compliance Management Workflows:** Automated generation of audit-ready records suitable for regulatory environmental (EPA Quad-O) or safety submissions. * **Self-Hosted Air-Gapped Deployment:** 100% offline, self-contained container-based deployment package running local LLM inference (Llama-3-70B or Mistral-Large) without external callbacks. ## Milestone M6: Repeat Visit (Q1 2027) Q1 2027 moves the platform from one-off campaigns to recurring data, and gives the assistant a reason to speak first: * **KRSI Robot Ingestion Adapter:** Robot and patrol captures enter through the existing ingestion path, keeping the platform hardware-agnostic and requiring no site infrastructure. * **Assistant-Initiated Observations:** Scheduled passes over newly landed data compose a per-engineer digest — the durable-execution shape rather than chat, since the result must outlive the run that produced it. * **Attention Ranking:** Assets ordered by attention debt (unreviewed evidence × cross-source and design corroboration × time since last review), each factor shown rather than collapsed into one opaque score. Presumes the evidence-confidence gates delivered in M5. * **Review-Throughput Surface:** Notes before findings, per-campaign worklist with generated reports, and reversible dismissal. ## Milestone M7: Continuous Coverage (Q2 2027) Q2 2027 instruments the site itself, so coverage is orchestrated rather than planned by hand: * **Navigation Beacons & Communication Backbone:** Fixed infrastructure lifting continuous coverage past ~90-minute battery endurance toward the ≥ 20 h/day objective. * **Full Spatial Navigation:** Autonomous traversal against the persistent facility model. * **Coverage Orchestration:** Next-best-inspection computed from staleness and risk across the asset register. * **Proactive Coverage Prompts:** Orchestration outputs surfaced as recommendations — never as scheduled actions. * **Fleet Intelligence Analytics:** Route history, endurance and coverage rolled up across the fleet. * **Self-Hosted Air-Gapped Deployment:** Offline container package with local inference. ### Q4 2026 — Research-Dependent Spikes These features are contingent on feasibility spikes scheduled in parallel: * **Physical AI Remediation Advice:** Grounding generative repair and maintenance advice inside the engineering domain model under safety-critical reliability thresholds. * **Facility-Specific Generative Training:** Calibrating synthetic thermal/OGI training data to the precise spatial profile of the target facility to bypass physical data scarcity. --- ## Long-Term Horizon (H1 2027 and Beyond) These capabilities are directional and extend beyond the committed 2026 milestones (M4–M5), representing our future product vision: * **Autonomous Patrol & Navigation:** Fixed-infrastructure robotics — navigation beacons, communication backbone, coverage orchestration, and fleet analytics — with full spatial navigation, for repeatable robot-patrol continuity decoupled from onboard SLAM. * **Dose-Aware Operations:** High-hazard environments extension to integrate radiation dosimetry and path planning through high-dose zones, where avoiding human entry is itself a primary economic and safety driver. * **Regulated-Environment Deployment Profile:** Lifetime data retention configurations, high-security on-premise profiles, and automated compliance packaging suitable for highly regulated industrial sites.