Runner Engineering · GSI
Otherwise known as Artificial Intelligence (AI)
Agents and information extraction, engineered like any other system that has to work: specified, measured, permissioned, and reviewed. GSI assembles models, tools, data, workflows, memory, evaluation, and human controls into one accountable system.
A plain definition
The intelligence is generated: composed from parts that can each be inspected, tested, and replaced. That is what makes it dependable enough to put into real engineering, quality, and operations work.
Reference architecture
Every GSI system is tailored, but the layers are consistent. Each one has its own tests, its own failure handling, and a clear contract with the next.
Connectors, format conversion, de-duplication, and access-scoped storage so every later step knows where a fact came from.
OCR, layout and table parsing, then schema-first extraction of fields, entities, and relationships with source spans attached.
A tiered cascade—deterministic rules, embedding models, then a language model—with confidence thresholds that decide what proceeds automatically.
Hybrid search and retrieval, knowledge graphs where relationships matter, and memory with explicit scope and retention.
State-machine and graph agents, planner–executor loops, and event-driven workflows calling permissioned tools in a sandbox.
Information extraction
Extraction is rarely one model. It is a pipeline of complementary techniques, each validated against the documents and signals you actually have.
Bring documents, scanned forms, drawings, photos, audio, sensor streams, and enterprise system records into one governed pipeline.
Each input keeps its origin, timestamp, and access scope, so downstream answers can be traced and permissions are never widened by accident.
Recover reading order, headers, title blocks, multi-page tables, and checkboxes—not just a flat wall of text.
Layout-aware parsing is selected and tested against your real documents, including poor scans and revision-marked drawings.
Define the target record first—fields, types, units, allowed values—then extract into it and validate before anything is stored.
Field-level confidence and source spans travel with every value, so a reviewer sees exactly where a number came from.
Identify parts, suppliers, assets, failure modes, people, and dates, then resolve them against your master data.
Normalization and entity resolution turn "Brkt, LH" and "left-hand bracket" into the same part record.
Hybrid keyword and vector search, reranking, and permission-aware retrieval so answers are grounded in your sources.
Responses cite the passages they rely on; when the sources do not support an answer, the system says so.
Model relationships explicitly when questions depend on them: part-to-assembly, failure-to-cause, requirement-to-test.
Used when multi-hop questions or traceability demand it, not by default; many problems are better served by search alone.
Classification architecture
Deterministic first, learned next, language models last. Each item stops at the cheapest tier that can classify it with enough confidence—and anything uncertain goes to a person.
Accept
Confidence above a threshold set from measured results proceeds automatically and is logged.
Escalate
Uncertain items fall through to the next tier or a second method for agreement checking.
Human review
Low-confidence or high-consequence items go to a review queue; decisions feed back as new training and test data.
Tier 1
Confidence gate ↓
Patterns, lookups, code lists, and business rules. Transparent, fast, and exact where the answer is already known.
Typical fit: Part numbers, document types with fixed signatures, regulated code lists.
Tier 2
Confidence gate ↓
Embedding similarity and compact supervised models trained on your labeled examples. Inexpensive and consistent at volume.
Typical fit: Recurring categories with enough examples: defect codes, complaint types, routing queues.
Tier 3
Confidence gate ↓
A large language model with structured output for ambiguous, long-tail, or context-heavy cases the earlier tiers cannot settle.
Typical fit: Free-text narratives, novel failure descriptions, multi-document judgments.
Final gate
Reviewers see the evidence and each tier's vote. Their decisions become new labeled examples and golden-set cases.
Agent frameworks & patterns
We select the simplest agent pattern that meets the requirement, implement it on established open frameworks or plain code as the situation warrants, and engineer the guardrails alongside it.
Explicit states, transitions, and checkpoints. The model decides within a step; the graph decides what steps exist.
A planner proposes steps; an executor calls typed tools—search, calculators, databases, internal APIs—and checks results.
Agents triggered by a new document, a sensor threshold, or a ticket, running in queues with retries and idempotency.
Separate specialized agents for distinct roles such as extraction, verification, and drafting, coordinated by a supervisor.
Working memory for a task, durable memory for approved facts, and nothing retained that policy does not allow.
Agents draft; people approve. Consequential actions—sending, ordering, changing records—pause for a named reviewer.
Trust, security & deployment
The controls are part of the design from the first prototype, not added after a demo succeeds.
Least-privilege credentials, allowlisted tools, isolated execution for generated code, and network egress limited to what the task needs.
Every extracted value and generated answer links back to its source document, page, and region.
Representative test sets built with your experts, scored before release and on every model, prompt, or data change.
Traces of each step, tool call, cost, and latency, plus dashboards for accuracy drift and review-queue volume.
Prompt-injection defenses for untrusted content, secrets kept out of prompts, and clear rules on which data may reach which model.
Versioned prompts, models, schemas, and thresholds with rollback, so behavior changes are deliberate and reviewable.
Placement follows latency, connectivity, data sensitivity, model size, and cost—and a single system can span several targets.
Edge device
On-device or gateway inference where latency, connectivity, or data locality rule out a round trip.
On-premises
Models and data inside your facility network for sensitive engineering or production data.
Private cloud
Your cloud tenancy with managed scaling, private networking, and your own keys.
Managed model APIs
Hosted frontier models where policy allows, behind a gateway that controls what is sent.
Representative applications
Illustrative patterns drawn from the problems Runner Engineering already works on—field quality, production, supply chain, and technical knowledge. Each is scoped to your data and constraints during discovery.
Classify claim narratives and technician notes into failure modes, link them to parts and builds, and route emerging clusters to engineering.
Pull part numbers, revisions, materials, and tolerances from drawings, change notices, and specifications into validated records.
Check certificates, inspection reports, and approval packets against requirements and flag what is missing or out of specification.
When telemetry crosses a threshold, retrieve the relevant manuals and history, then draft a work order for a technician to approve.
Answer questions across standards, procedures, and lessons learned with citations, scoped to what each user is permitted to see.
Summarize shift logs and nonconformance reports, classify deviations, and assemble evidence for disposition reviews.
Engagement path
Each phase ends with a decision point. You can stop after any phase with something useful in hand.
Define the decision or workflow, the users, data sources, failure consequences, and what success must be measured against.
Leaves behind: Problem statement, data inventory, risk register, and acceptance criteria.
Test the riskiest assumptions on your real samples and build the first golden set with your subject-matter experts.
Leaves behind: Working prototype, baseline evaluation results, and a go / no-go recommendation.
Integrate with real systems, run in shadow mode or with full human review, and tune thresholds from observed results.
Leaves behind: Integrated pilot, review-queue tooling, and measured performance on live inputs.
Security review, permissions, monitoring, rollback, runbooks, and deployment to the chosen edge, on-premises, or cloud target.
Leaves behind: Production system, operating documentation, and an evaluation suite that runs on every change.
Watch drift and review volume, refresh test sets, and evaluate new models against your own benchmarks before adopting them.
Leaves behind: Ongoing performance reports and a controlled improvement backlog.
Questions
No. GSI is our name for an engineering discipline: assembling dependable intelligence from models, tools, data, workflows, memory, evaluation, and human controls. Otherwise known as artificial intelligence (AI)—engineered to be measured and accountable.
No. Many extraction and classification tasks are best served by rules, layout parsing, or compact trained models. Language models are used where they measurably outperform simpler methods, and the cascade keeps them off the cases that do not need them.
Often, yes. Depending on the task and model size, systems can run on edge hardware, on-premises servers, or a private cloud tenancy. The deployment target is chosen from your data, latency, and policy constraints.
By measuring it. We build golden test sets with your experts, agree on acceptance criteria before production, and re-run the evaluation whenever a model, prompt, schema, or threshold changes.
Custom AI engineering covers the full hardware-to-software system: sensors, edge compute, data, models, and applications. GSI goes deeper into the intelligence layer—agents, extraction, classification, and the controls around them—and often sits inside that larger system.
Start with one workflow
Share samples, the decision it supports, and what a wrong answer would cost. We will help define what can be automated, what should stay with people, and how to prove the difference.