New · Forge Suite · v1.0

ForgeMind

The GIS Intent Compiler

Natural language in. Correct geospatial pipelines out. Or an honest refusal. ForgeMind is the natural-language control plane for ForgeGIS — it turns a spoken or typed request into a validated, executed pipeline on the GPU-accelerated engine, and when a request cannot be honored correctly it returns an explicit refusal rather than a confident guess.

Zero
confident-wrong results across 79 conversations and 6,696 turns
~25×
lower LLM cost vs. an agent-loop baseline on the same workflows
97
voice-routable capabilities exercised end to end
~94%
compiler eval pass rate: intent → validated pipeline

What it is, why it matters, how it's built differently

Three short answers for the technical evaluator.

What it is

ForgeMind sits between a human (or another agent) and the ForgeGIS compute engine. It decides what to run, binds the request to real data, dispatches the work, and narrates the result — turning informal language into a correct, reproducible operation. It speaks the Model Context Protocol natively in both directions.

Why it matters

A conventional agent lets a model author tool chains directly, so every wrong parameter and hallucinated step lands in production looking confident. ForgeMind refuses by construction: a request either compiles to a validated pipeline or is declined. That is what regulated and mission work requires.

How it's built differently

The language model classifies intent and extracts parameters only. Deterministic code owns pipeline structure, type-checking, and execution — so the same request yields the same pipeline every time, with no hallucinated params and no drift. Cheap work runs on a fast model tier; the costly tier is reserved for the hard step.

The compiler model

ForgeMind treats a natural-language request the way a compiler treats source code. The language model does one narrow job — extract intent into a typed intermediate representation. From there, deterministic machinery takes over.

Natural language
“Compute least-cost path from ingress_ip to hlz_charlie”
LLM: intent extract
Maps words to a typed intent. No execution.
Typed IR
Operation + resolved anchors + parameters
Deterministic check + route
Type-check, resolve refs, refuse if unsound
GPU execution
ForgeGIS runs the validated operation
Typical LLM agent loop ForgeMind compiler
What the LLM does Plans and executes, reasons about geometry directly Extracts intent into a typed IR, nothing more
Where math happens Inside the model's reasoning (unverifiable) Deterministic engine · GPU-accelerated ForgeGIS
Failure mode Plausible, confident, wrong Explicit refusal. Never a fabricated number
Cost / latency Many model calls per task One intent pass, then deterministic execution
Auditability Opaque chain of thought Inspectable IR for every single turn

Architecture commitments

The internal GIS Intent Compiler is proprietary; what follows is the externally observable contract each commitment produces.

Refusal by construction

Validation before execution

Every candidate pipeline is checked against a type system and operational guards before any work is dispatched. A request that cannot be mapped to a valid operation, lacks its parameters, or cannot bind to data is refused with a specific, machine-readable reason — never run speculatively.

Narrow the model

The LLM does only intent

The language model is confined to classifying what the operator wants and extracting parameters. The shape of the pipeline is owned by deterministic machinery, so the same request produces the same pipeline every time — no hallucinated parameters, no operation-arity slips, no run-to-run drift.

Cost engineering

The right model for each job

Cheap, high-volume classification and narration run on a fast, low-cost tier; the more capable tier is reserved for the genuinely hard reasoning. The measured effect is roughly a 25× reduction in language-model cost versus an agent-loop approach — without sacrificing correctness.

No vendor lock-in

LLM-agnostic by design

The provider and model sit behind one abstraction — configuration, not architecture. Switch models to follow cost, capability, availability, or procurement constraints (including continuity during a provider outage) without code changes. Claude and OpenAI ship built in; self-hosted models use the same integration point.

Learns with use

Semantic memory

Successful workflows, and corrections to unsuccessful ones, accumulate as retrievable exemplars backed by local storage and on-device embeddings. Learning happens through accumulated retrieval context, not model fine-tuning — so it is inspectable, clearable, and never silently bakes a bad pattern into a model's weights.

MCP citizen

Native MCP, both directions

As a client it dispatches to MCP tool servers — the ForgeGIS compute surface, a dataset catalog, a map client — guarding each namespace behind a circuit breaker. As a server it exposes a single high-level tool a higher-order agent can hand an entire geospatial workflow to.

Geometry the agent can reason about

Named scene entities

Named geometry a user places on a map — a dock, a route, an objective — is referenced by a stable handle and hydrated from the catalog on demand, never re-derived from a bounding box. Fail-closed: if the true geometry is not reachable, the operation defers and asks rather than emitting a plausible-but-wrong product.

Showcase · Session S-MARATHON-22

One conversation. A complete mission plan.

A 102-operation test scenario simulating a combat search-and-rescue plan, run end to end in plain English against real SRTM terrain. Every operation was understood, type-checked, routed, and executed — and every one succeeded.

102
operations, one session
29
named entities, all reused
27
distinct operations
0
failed or confident-wrong
Read the full marathon showcase →

Capabilities, by intent

ForgeMind makes the 239-operation ForgeGIS catalog reachable in natural language — anchored on a deterministic core of more than 130 routed capabilities, with 97 exercised end to end in the validation campaign. Coverage grows capability by capability as each is validated; ForgeMind advertises only what it can run correctly.

Direct analysis. “Compute viewshed from 34.5N 69.2E with a 10 km radius.” — explicit-coordinate raster analysis.
Discovery into analysis. “What elevation datasets do we have for this region? Pick the highest-resolution one and run a slope analysis.” — a catalog query chained into a compute operation.
Site suitability. “Where can we land outside the enemy viewshed?” — a multi-operation suitability workflow combining line-of-sight exclusion with terrain constraints.
Scene-aware. “Mark Objective Bravo here and map the slope around it,” then later “rebuild the slope picture — it displaced to a new grid reference” — named-anchor placement and refinement across turns.
Visualization. “Load the result onto the globe as a red, 50% transparent layer.” — chaining a compute product into a map-client display.

Validation and coverage

ForgeMind is validated along two complementary axes — a large multi-turn conversation campaign that measures end-to-end behavior as an operator would experience it, and a focused compiler-evaluation campaign that measures the core intent-to-pipeline transformation in isolation.

Full validation methodology and campaign detail: ForgeMind Technical Brief (PDF).

Deployment and integration

ForgeMind ships as a single self-contained Java application, deployable where the mission runs.

Two shapes, one artifact

Run it as an HTTP service exposing a small REST surface — a buffered endpoint plus a streaming (SSE) endpoint that surfaces each model round trip and tool call live — or as an MCP stdio server that presents ForgeMind as a single delegatable tool to a higher-level agent host. Auth, rate limiting, request-size caps, and a concurrency cap are built in.

Purpose-built for ForgeGIS

ForgeMind spawns and supervises its MCP backends — the ForgeGIS compute engine, a dataset catalog, and a map client — as local subprocesses over stdio, with health checks and per-namespace circuit breakers. ForgeGIS is the engine it is built to drive; the catalog and map-client roles are pluggable behind the protocol.

Fit for regulated & air-gapped sites

A pure-JVM application with local-disk persistence and no required external services beyond the language-provider endpoint. The memory store is a single local database file; backends run over stdio, not the network. For programs hosting their own models, the provider abstraction is the single integration point.

Who it's for

One product, three readers. Each one-pager leads with the parts of ForgeMind that matter most to that audience.

Downloads

All ForgeMind collateral. Direct download — no email gate, no form wall.

Talk to us about ForgeMind

Licensing questions, evaluation access, partnership and distribution inquiries — we read every email. License terms are being finalized; an evaluation build is available to qualified teams on request.

rich@seaglassfoundry.com