Glaux

Enterprise business processes are broken in four different ways, each solved by a different category of tool:

• Process Mining (Celonis, UiPath) — discovers how processes actually run, but these tools are expensive, single-LLM, and require consultants • Regulatory Compliance (GRC platforms) — track regulations but don't ingest documents, extract controls, or map them to processes • Remediation (RPA tools) — automate tasks but don't do root-cause analysis or agentic decision-making • KPI Tracking (BI dashboards) — visualize metrics but don't connect them back to process behavior

The result: enterprises pay $M+/year for 4 separate tools that don't talk to each other, plus consultants to bridge them. The thesis behind Glaux is that these four capabilities should be one platform because the data flows between them — process mining identifies where KPIs are missed, the regulatory engine maps where compliance is at risk, and the remediation engine closes the loop with automated fixes.

There's also a deeper market gap: 80% of global enterprises still run on legacy infrastructure and are excluded from the AI revolution because every AI solution assumes Silicon Valley-level resources. Glaux is designed for them — on-prem-capable, data-residency-aware, and built to work with what enterprises have today, not what they might have.

Building four enterprise-grade engines that typically ship as separate products, then making them work as one system.

Process Mining engine — 10 discovery algorithms (Alpha, Heuristics, Inductive, DFG, ILP, POWL, Correlation, DECLARE, Log Skeleton, Temporal Profile), 12 conformance techniques (Token Replay, Alignments, Anti-Alignments, Multi-Alignments, Decomposed Alignments, Process Tree Alignments, DFG Alignments, Edit Distance, Footprints, Temporal Profile, DECLARE, Log Skeleton), real-time streaming, 3 ML prediction models (next activity, remaining time, case outcome), 3 anomaly detection methods (Isolation Forest, One-Class SVM, DBSCAN). LLM-generated process maps from natural-language descriptions, deviation analysis that explains why (likely cause, business impact, recommendation), and natural-language Q&A against process data.

LLM-native architecture — every process mining output (discovery, conformance, performance, deviations) is designed to flow through an LLM for explanation, not presented as raw metrics. Process data is vector-indexed on ingestion so a RAG-powered chatbot can answer questions against specific cases, variants, and deviations with grounded citations.

Regulatory engine — Claude-powered PDF ingestion with 128K context window, control extraction from dense legal documents, live regulatory feed subscription across jurisdictions, process-to-regulation mapping.

Remediation engine — Agentic AI with JSON-structured recommendations, ROI estimation, automated action execution, RCA via ML + LLM.

Multi-LLM architecture — Runtime switching between Claude (Sonnet 4.6 default + Opus 4.7 for deep reasoning) on Anthropic, GPT 5.4 on OpenAI, and Llama 3.2 via Ollama (on-prem). Essential for enterprise data-residency requirements, not optional.

Scale handling — Event logs larger than any single LLM context window are handled via chunked multi-turn conversations with the model, with truncated JSON continuation for incomplete responses. Heavy analysis runs asynchronously via Celery background workers.

Data infrastructure — 5 data services (MongoDB for users/audit/event logs/process models, TimescaleDB for time-series performance metrics, Neo4j for graph queries, Redis for caching + pub/sub, Kafka for real-time streaming), 47+ APIs, WebSocket streaming for real-time updates.

All of this needs to work as one cohesive product, not four integrations bolted together.

Built the platform as two integrated services: a React 18 + Vite + TypeScript frontend with ReactFlow for process visualization, Tremor for dashboards, Material-UI for admin panels, Zustand for state, and React Query for server state. The backend is two services working together — an Express.js gateway for auth, admin, and the regulatory engine, plus a FastAPI + Python service for process mining (built on PM4PY) and ML pipelines.

Process Mining engine: Ingests event logs in CSV/XES/XLSX/Parquet → runs PM4PY's 10 discovery algorithms → produces BPMN visualizations via ReactFlow + Dagre layout. The LLM-powered generator can also create process maps from natural-language descriptions — a user describes a process in plain English, and Claude generates the map. The deviation analyzer doesn't just flag anomalies; it explains the likely cause, business impact, and recommended action for each one. ML pipeline uses scikit-learn for predictions and anomaly detection. Results flow through Claude for natural-language explanations with context, factors, and confidence levels. Celery handles heavy analysis as async background jobs so the UI stays responsive.

Vector-indexed RAG chatbot: On event log ingestion, the system vector-indexes cases, variants, activities, and deviations into a searchable knowledge base. Users ask questions in plain English — "why did case #4521 fail?", "what's slowing down Q3 orders?", "which regulations apply to this process?" — and the chatbot retrieves relevant process artifacts, feeds them to the LLM, and answers with grounded citations. This turns process mining from a consultant-operated tool into something business analysts can actually use.

Regulatory engine: Claude-powered PDF chunking handles documents of any size (40-page chunks into 128K context), extracts controls with responsible parties, deadlines, and priorities, tracks live regulatory feeds across jurisdictions, and maps processes to regulatory requirements.

Remediation engine: Agentic AI analyzes deviations and anomalies, generates JSON-structured recommendations with ROI estimates and implementation steps, routes them to approval workflows before execution.

Multi-LLM: Provider abstraction layer with runtime switching via database settings. Organizations with data-residency requirements run Ollama locally; others use Claude or GPT. Prompt engineering is tuned per provider for optimal output.

Scale handling: Event logs larger than any single LLM context window are processed via chunked multi-turn conversations with the model, with truncated JSON continuation handling incomplete responses gracefully — a practical necessity for enterprise-scale logs.

Real-time: Socket.IO for activity logs, dashboard updates, and process mining streams. WebSocket endpoints for live case replay and conformance visualization.

Unified platform delivering what previously required four separate enterprise tools — process mining, GRC, RPA, and BI — into one system with shared data and shared intelligence.

Advisory validation: Ex-COO of HP Enterprise chose to advise Glaux as his only post-retirement commitment. His advisory decision was based on the thesis that 80% of global enterprises running legacy infrastructure are excluded from the AI revolution — and that the real market is serving them, not Silicon Valley.

Pilot traction: Advancing pilot discussions with 2 Fortune 500 enterprises. VC conversations framed around the $10B+ process intelligence market.

Technical depth: 47+ APIs across two services, 10 discovery algorithms, 12 conformance techniques, 3 ML prediction models, 3 anomaly detection methods, 5 data services, 3 LLM providers, real-time streaming engine, Claude integration with 128K context and extended thinking, natural-language process modeling.

Hybrid statistics + AI delivers far more reliable results than pure LLM inference. Let the math handle structure (100% coverage of transitions, variants, gateways, loops), and let AI handle insight (pattern recognition, natural language explanations). Pure-LLM process mining tools sample data and miss edge cases.

Multi-LLM isn't a feature — it's an enterprise requirement. Organizations with data-residency constraints need Ollama (local), while others want the best cloud model. Runtime switching via database settings turns this from a deployment blocker into a non-issue.

LLM-native beats LLM-bolted-on. Most incumbents (Celonis, UiPath) added AI features on top of existing products — chat widgets, summary buttons, explanation tooltips. Building LLM-first means every output flows through the model for context, every query is natural-language, every deviation is explained rather than coded. The architecture is fundamentally different, and users feel it immediately.

Natural-language process modeling is the unlock for business-user adoption. The gap between "process mining tool" and "tool a business analyst actually uses" is exactly this — being able to describe a process in English and get a map. Most process mining tools are consultant-operated; making them business-user-friendly through LLM interaction is what opens the market beyond F500 early adopters.

The four-engine thesis was correct: process mining, regulatory compliance, remediation, and KPI tracking belong together because the data flows between them. Treating them as separate products is how the industry organized historically, not how the problem actually maps to the data.

On-prem-first architecture forces discipline. Every component must be self-contained and operationally simple — you can't rely on managed services. This is harder to build but produces a fundamentally more deployable product for the F500 enterprises running legacy infrastructure.