Applications And Market Map¶
Director-AI is a factual-coherence control plane for LLM applications. It sits between generated output and consequence: a user-visible answer, a streamed token path, a tool call, an agent handoff, a stored record, or an audit event.
Use this page when you need to explain the product quickly to a buyer, developer, evaluator, operator, or partner.
What The Software Is¶
Director-AI is not a chatbot, prompt template, content-moderation list, or model provider. It is a guardrail runtime that checks whether generated text is grounded enough to proceed.
The runtime combines:
- governed facts from inline rules, document ingestion, or vector retrieval;
- configurable scorers, including rules, embeddings, NLI, and structured verification;
- response-level factual-coherence scoring, with opt-in contradiction-driven streaming halt for completed streamed claims;
- SDK, framework, REST, gRPC, inference-server, and voice integration surfaces;
- tenant-safe evidence, audit, metrics, and forensics records.
The smallest useful path is director-ai-lite, which exposes a three-line
guard facade for free-tier adoption. The full director-ai package is
the open-core runtime. Director-Class AI is the commercial implementation and
evidence programme around governed customer deployments; it is not a separate
wheel.
Who Needs It¶
| Audience | Primary problem | Director-AI value |
|---|---|---|
| Product owner | A wrong answer damages trust or triggers rework | Define which answers need fact gates and what evidence a pilot must collect |
| Application developer | LLM output reaches users before review | Add guard(), SDK middleware, or a REST proxy at the output boundary |
| RAG engineer | Retrieval context is stale, noisy, or incomplete | Score answers against governed facts and retain retrieval evidence |
| Runtime/platform team | Multiple apps need one shared control layer | Deploy REST/gRPC, auth, metrics, audit, and rollout controls once |
| Evaluation lead | Model, prompt, or KB changes need regression gates | Run batch scoring, thresholds, false-positive review, and benchmark cards |
| Governance/security team | Incidents need reviewable evidence without raw data exposure | Use tenant-safe events, compliance reports, and guardrail forensics |
Application Lanes¶
Pick one lane for the first pilot. Each lane has a different success signal.
| Lane | Protected workflow | First proof |
|---|---|---|
| Customer support | Refunds, warranty, policy, entitlement, and account answers | One known-good answer approved and one unsupported policy answer rejected |
| Enterprise knowledge assistant | Private-document answers and summaries | Verdict includes retrieved chunks and a traceable rejection reason |
| Regulated review | Medical, legal, finance, or research drafts | Unsupported claims route to human review with evidence and threshold rationale |
| Streaming assistant | User-visible token stream | Contradictory completed claim halts before the stream continues |
| Agent workflow | Tool output, chain step, or handoff | Unsafe step rejects or routes before downstream action |
| Evaluation pipeline | Prompt/response datasets and model updates | Batch report shows threshold, false positives, and false negatives |
| Platform deployment | Shared guardrail service | Auth, metrics, logs, rollback, and runbook evidence are visible |
Market Value¶
The market value is control over factual risk. Director-AI can reduce:
- unsupported customer-facing claims;
- manual review load for routine factual checks;
- regressions from model, prompt, or knowledge-base changes;
- incidents caused by hallucinated streamed output;
- duplicated guardrail work across LLM providers and application teams.
It can increase:
- buyer confidence that LLM output remains tied to governed facts;
- operational visibility into why an answer was allowed, rejected, halted, or routed;
- reuse of one guard policy across SDK, REST, gRPC, voice, agent, and inference-server deployments;
- quality of procurement, security, and compliance evidence.
Director-AI does not remove the need for domain experts, governance, access control, or legal review. It gives those functions a concrete enforcement point inside LLM output flow.
What Ships Publicly¶
| Surface | Public package/docs | Commercial extension boundary |
|---|---|---|
| Director-Lite | director-ai-lite, three-line guard facade, free onboarding |
None required for free-tier use |
| Director-AI | director-ai, open-core SDK, scorers, APIs, integrations, evidence packet, docs |
Paid Pro use for Advanced & Labs production deployment |
| Director-Class AI | Promoted in docs and PyPI positioning | Customer-specific deployment, sector packs, evidence reviews, tuning, and SLA |
The public repository contains the core software and general evidence surfaces. Customer-specific sector packs, tuning datasets, deployment recipes, acceptance criteria, and performance claims must be validated against the customer's own governed data before they are used commercially.
First Evidence Packet¶
Before any serious pilot discussion, produce a small evidence packet:
pip install "director-ai[nli]"
director-ai evidence --emit evidence/
director-ai verify-evidence evidence/
That packet should prove:
- a governed fact loaded successfully;
- a grounded answer passed;
- a hallucinated or contradictory answer failed;
- the decision record has a digest and can be verified;
- the operator can explain what happened without exposing secrets.
Documentation Path¶
| Need | Read next |
|---|---|
| Product and tier boundary | Product Overview |
| Buyer value and budget language | Market Value and Positioning |
| First local run | Quickstart |
| Guided notebook path | Notebook Gallery |
| API selection | API Reference |
| Pilot checklist | Evaluation Onboarding |
| Production operation | Production Guide |