Our approach to AI

AI that helps.
Never AI that decides.

TraceBoard's engineering data — requirements, links, test results, baselines — is always deterministic and always yours. AI is an optional layer on top of that core: it can propose, draft, and flag, but nothing enters your system of record without your explicit approval. The product works fully offline with zero AI configured, and it always will.

Every AI-assisted change follows the same path
PROPOSE AI drafts a suggestion VALIDATE Checked against schema APPROVE A person decides COMMIT Deterministic. Final.
The problem

Compliance data can't be "probably right."

Most ALM vendors handle AI one of two ways, and neither works for regulated teams.

Bolted-on AI

Legacy tools add a chat box or an "AI assist" button as a roadmap checkbox, with no architectural line between what the model suggests and what the system treats as fact. It looks modern in a demo and becomes a liability in an audit.

AI-native tools

Newer tools build the model into the core workflow itself — convenient, until a compliance engineer has to explain to an auditor why a traceability link exists because "the model said so." For a system of record, that's not an acceptable answer.

Principles

Four rules we don't bend on

01

Deterministic core, always

Requirements, links, and test data are structured facts placed there by a person — never model output treated as ground truth.

02

Approval is mandatory

Nothing an AI feature proposes becomes part of your record until a human reviews and accepts it. No silent writes, ever.

03

Optional at every level

From a single suggestion to the entire deployment, AI can be switched off with no loss of core functionality — not a degraded mode, the same product.

04

Works fully offline

Self-hosted and air-gapped deployments run with zero AI, zero cloud calls, and zero telemetry. This isn't a fallback — it's the baseline.

How it works

Propose → Validate → Approve → Commit

The same four-stage pipeline governs every AI feature in TraceBoard, regardless of which module it touches. Take test case generation as an example.

01 · PROPOSE

AI drafts test cases from a requirement

Given a requirement's text and context, the model drafts candidate test cases — nothing is written to the project yet.

02 · VALIDATE

Structure is checked deterministically

Drafts are checked against your schema — required fields, valid states, correct linkage — before a person ever sees them.

03 · APPROVE

You review, edit, or reject

Every draft is shown clearly as a suggestion. Accept it, edit it, or discard it — the choice is always visible and always yours.

04 · COMMIT

Only approved items enter the record

Once approved, the test case becomes a normal, deterministic TraceBoard object — indistinguishable in the system from one written by hand, with a full audit trail of who approved it and when.

Where AI helps today

Real workflows, not a chat window

Every item below only proposes. Nothing here writes to your project without approval.

PROPOSES

Trace analysis

Surfaces coverage gaps across requirements, tests, and documents, and helps you prioritize which ones actually matter.

PROPOSES

Messy Excel & Word cleanup

Bring in inconsistent spreadsheets or documents as they are. AI proposes a clean, mapped structure — you approve it before anything imports.

PROPOSES

Test case generation & analysis

Drafts test cases from requirements and flags existing tests that may be stale after a requirement changes.

PROPOSES

Document generation & analysis

Drafts report sections from your project data and flags inconsistencies across your document set.

ON THE ROADMAP ReqIF schema mapping on import, and project-grounded natural language search — same pipeline, same approval gate.
What never changes

The system of record stays deterministic

COMMITTED

Requirements, links & baselines

Structured, typed data — never generated on the fly by a model.

COMMITTED

Audit trail

Every change traces back to a specific human decision, including approved AI suggestions.

COMMITTED

Compliance data

Coverage, sign-offs, and traceability reports reflect only what was approved — nothing pending, nothing inferred.

COMMITTED

Full offline operation

The entire product, including this list, works with zero AI configured.

Built for the environments that can't take chances.

TraceBoard deploys self-hosted, including fully air-gapped. AI is one optional Docker service among several — remove it, and every other module runs exactly the same.

NO CLOUD DEPENDENCY
Core product has no external service calls
NO DATA LEAVES YOUR INFRASTRUCTURE
Local model support for fully offline AI
NO AI REQUIREMENT, ANYWHERE
Every module works with AI fully disabled
Model compatibility & data freedom

Your models. Your data.

TraceBoard Suite adapts to virtually any LLM and keeps your compliance data portable — no proprietary lock-in.

Broad LLM Compatibility

The TraceBoard Suite is engineered to work with virtually any Large Language Model. Whether you require a tiny 3B parameter model running locally on a low-power device, or a 1.6 trillion parameter model from a cloud provider, the architecture adapts seamlessly.

  • Local (air-gapped): supports Ollama, LM Studio, and any OpenAI-compatible local server.
  • Cloud: natively supports Mistral AI, OpenAI, Anthropic, DeepSeek, and any provider with an OpenAI-compatible API endpoint.

Smaller models (≤7B) may produce less detailed narratives, while larger models provide deeper insights — giving you full control over the trade-off between cost, latency, and quality.

Commitment to Data Freedom

At TraceBoard, we believe your compliance data is yours, not ours. We explicitly do not use proprietary binary formats or encrypted data stores.

  • Open Standards: all import/export relies on JSON, PDF, SQL, and DOCX.
  • Full Database Access: you have full access to your PostgreSQL database.
  • Zero Lock-In: you are free to export your entire project — requirements, tests, tasks, and traceability links — at any time, without requiring our assistance.

See where the line actually sits.

The best way to understand the approval gate is to watch it work on your own data — a messy spreadsheet, a real requirement set, or a document you're tired of formatting by hand.