Signals solves the “Enterprise Data Tax” on strategic thinking.

Example: In commodity-exposed industries, analysts manually spend hours extracting data, normalising units, refreshing pivot tables, and writing comments before strategic analysis can even start. Signals automates this bridge, synthesizing internal BI, CRM, and ERP parameters with external market data in under 30 seconds with zero manual friction.

Traditional BI systems are built to report on historical internal data that has already been cleaned and structured. They do not bridge the gap to live, dynamic external market signals (e.g., crude pricing shifts affecting MDI feedstock costs with a 2–4 week lag).

Signals feeds your BI—it doesn’t replace it. It outputs clean, pivot-ready Excel workbooks directly into your Power BI inputs, and lets you drag and drop database exports to get instant market-contextualized reports.

An Avatar is a purpose-tuned vertical instance of the Signals platform configured for a specific industry’s inputs, feedstock chains, and decision criteria.

For example, ChemSignals (Avatar 1) is calibrated specifically for the petrochemical value chain, tracking crude derivatives, natural gas, feedstocks like benzene/aniline, and macroeconomic parameters to output weekly intelligence briefs and margin indicators.

The platform architecture is vertical-agnostic and generalizes to any sector where margins are driven by live inputs that fluctuate faster than the internal reporting cycle. Examples:

• PharmaSignals: supply chains, contract manufacturing, FDA regulatory updates.
• AgriSignals: Grain, oilseed, and livestock feedstocks, CBOT futures, weather metrics.
• EnergySignals: Power spot prices, gas storage levels, renewable capacity, grid congestion.
• MetalSignals: Steel, aluminum, rare earth inputs, LME prices, tariff policies.

In feedstock chains with a 2–4 week transmission delay (like crude-to-MDI), margin compression occurs silently before pricing renegotiations. A 1% margin compression from an undetected feedstock shock can cost a major chemical manufacturer millions of dollars in a single quarter. Signals detects these movements upstream, enabling proactive surcharge activation.

The core of the architecture is an internal Small Language Model (iSLM) in the 7B–13B parameter range (e.g., Llama 3.1 8B, Mistral 7B). Rather than generating text directly, the iSLM acts as a deterministic router at temperature 0.0, classifying queries into four routing pathways:

• RAG: routes queries to local static document databases.
• MCP: queries live external APIs and databases via the Model Context Protocol.
• hybrid: chains retrieval pipelines in sequence.
• direct: handles queries directly from local context.

We enforce a strict policy of Data Locality.

The iSLM, RAG vector database, and MCP agents run entirely within the enterprise’s private cloud VPC or local infrastructure. No proprietary document chunks, procurement costs, or database parameters are sent to external, shared AI APIs. All data reasoning remains contained within the corporate boundary.

MCP provides a unified, open-standard connector layer. Instead of writing custom API wrappers for every source, MCP allows the platform’s orchestration layer to query live external market feeds (yfinance, FRED, RSS feeds) and internal systems (like SAP ERP exports) dynamically using the same model-accessible interface.

Hallucinations are eliminated by separating routing logic from synthesis:

1. The iSLM acts solely as a router and does not generate data.
2. The requested context is pulled deterministically from RAG (static docs) and MCP (live feeds).
3. The synthesis layer receives this verified context and formats the output, matching every metric back to its original source hash.

• Frontend: Next.js (React) using Recharts panels for data visualization.
• Orchestration: Python pipelines integrating `yfinance` and the FRED API.
• Integrations: SheetJS (`xlsx`) for generating pivot-ready bidirectional reports.
• Routing & Inference: Local small language models (7B–13B) run via containerized endpoints.