SENTINEL AI logo SENTINEL AI TAHAI Web Services

Local-first · governed · source-aware

SENTINEL AI

Governed AI that can work where your work lives.

A deploy-anywhere workbench for chat, files, legal research, coding, voice, browser-style workflows, and generated visuals — routed through Prefrontal Node so model choices, tools, memory, approvals, and evidence stay accountable.

Deploy anywherelocal · on-prem · cloud Governed intelligencepolicy · approvals · traces Source-aware workflowsfiles · citations · provenance
Watch the build See capabilities Open ecosystem Contact TAHAI
ChatVoiceFilesLegalCodingBrowserImageOperatorChatVoiceFilesLegal
governed chat interface
Ask SENTINEL to create a simple cat and display the result in chat.
SENTINEL
Image job routed through Prefrontal. The generated asset is returned inside the SENTINEL chat UI with provenance metadata.
Full SENTINEL AI governed chat interface showing a generated cat image displayed inline with provenance metadata.
Full chat proof Full SENTINEL AI workbench UI → governed route → generated image → provenance note → in-chat display.
route: image lanepolicy: recordedtrace: visible
Public proofVisible product progress without exposing internal build chatter.
Governed pathRouting, policy, sources, approvals, and traces stay central.
Local-first postureDesigned for private folders, local sources, and controlled deployment.
Open ecosystemOpen-weight models and OSS components are treated as strengths.

Honest launch path

Static site now. Commercial beta target: Summer 2027.

SENTINEL AI is being built in public. This site is the public product story and proof surface, not a hosted production app. The near-term focus is a reliable local Windows runtime, governed browser evidence, approved-folder file workflows, source-grounded answers, image provenance, and legal/coding review gates.

NowStatic public site, visual identity, architecture narrative, build proof, trust pages.
NextLocal runtime polish, SENTINEL Browser evidence, file/RAG workflows, image and legal proof.
TargetCommercial Beta in Summer 2027, with paid early access only after repeatable local proof.

Five-second understanding

SENTINEL is the workbench. Prefrontal is the control plane. The coding workbench is converging with both.

SENTINEL AI is the public platform experience: the place a user chats, attaches files, asks for legal research, requests a diagram, reviews a source, or approves an action. Prefrontal Node is the embeddable brain-sidecar underneath it: routing, policy, memory, tool control, approvals, and trace evidence. The adjacent coding app is becoming the build-and-verification lane in the same governed architecture.

SENTINEL AIUser workbench for serious AI work.
Prefrontal NodeGoverned orchestration and evidence layer.
Coding workbenchPatch, test, verify, and merge with review.

Deploy Anywhere

Run locally, on a workstation, on a private server, on-prem, in cloud, or in restricted environments.

Governed Intelligence

Every meaningful action can carry route, policy, source scope, approval state, and trace evidence.

Local Knowledge

Use approved folders, local documents, private sources, and reviewable retrieval without arbitrary device access.

Model Agnostic

Use local, hosted, open-weight, specialist, and challenger models through one governed routing layer.

Governed, not improvised

Prefrontal Node turns AI work into an inspectable route.

A request enters, gets classified, routes to the right lane, uses only allowed sources/tools, records evidence, and escalates sensitive actions for human review.

User requestPrompt, file, voice, image, task, or action intent.
Policy & routingRisk tier, permissions, model lane, tool lane, cost profile.
Memory / RAGApproved sources, chunk anchors, file metadata, citation status.
SENTINEL AI logoPrefrontal Nodeorchestrate · govern · trace
Model laneAssistant, legal, coding, vision, image, or challenger model.
Tool laneSearch, files, OCR, code, connectors, exports, generated assets.
Trace & evidenceRoute, source scope, policy decision, artifact metadata.
Human reviewApproval gates for legal, external, destructive, or high-risk actions.

System topology

Specialist lanes, one governed control plane.

Instead of forcing one model to do everything, SENTINEL can route work to specialist lanes while Prefrontal keeps the policy and evidence layer consistent.

AssistantGeneral reasoning, planning, summarization, and work coordination.
Files/RAGApproved-folder retrieval with source anchors and parser visibility.
LegalResearch and drafting support with source verification and review gates.
CodingRepo inspection, patch bundles, test evidence, and approval-first actions.
Image/VisionOCR, source-grounded visuals, diagrams, and generated asset provenance.
BrowserRendered web access, screenshots, DOM extraction, source cards, and governed search workflows.

More than a chatbot

A full AI workbench, lane by lane.

Every capability is presented as governed work: visible sources, clear route labels, approval gates, and reviewable outputs where they matter.

Chat & Voice

Natural conversation, push-to-talk voice, interruptible responses, transcripts, runtime labels, and action cards.

Open lane

Files & Local Corpus

Approved-folder workflows, large PDF handling, OCR, manifests, source anchors, split/rename review, and search.

Open lane

Legal Research

Matter organization, authority lookup, timelines, citation checks, quote verification, drafts, and review-required exports.

Open lane

Coding Agent

Repo inspection, patch proposals, test plans, sandboxed commands, diff review, and protected-write approval gates.

Open lane

Vision & Image

Scanned documents, screenshots, forms, image generation, editing workflows, asset metadata, and safety review.

Open lane

Governance

Policy, tenancy, approvals, trace evidence, source status, evals, release evidence, and model admission discipline.

Open lane

Governed Browser

Rendered web access, search-provider smoke checks, screenshots, DOM extraction, source cards, and policy-gated web research.

Open lane
SENTINEL workbench
Review this filing bundle, extract issues, and show what still needs verification.
SENTINEL
I found 42 source records, 8 timeline events, 5 authority checks, and 3 items that need review before export.
Action proposedGenerate review packet with source manifest, timeline, cite report, and draft checklist.

Primary product surface

A serious AI workbench, not just a text box.

SENTINEL is designed to show the work around the answer: source cards, route labels, policy state, approval cards, generated assets, and trace IDs. The user experience should feel powerful while still making the system accountable.

01Ask or attach
02Route through Prefrontal
03Ground to sources
04Review and export
Open interface preview
SENTINEL governed workspace
Review these materials and show me what is supported by sources.
Source scope selected. Retrieval and policy checks recorded.
I found three supported themes, two items that need verification, and one item that should stay marked review required.
model route visibletrace recordedreview gate
Evidence railFile, page, chunk, quote span, and status.
Action cardApprove, deny, or edit before high-risk work.
Generated assetImage, diagram, or export with provenance.

Local knowledge without unsafe access

Files stay user-approved, source anchored, and reviewable.

SENTINEL is designed to work from explicit folder permissions and attachment flows. It can scan, parse, OCR, index, retrieve, and cite from approved sources while leaving original files untouched.

Approve folderScan manifestParse / OCRIndex chunksRetrieve sourcesCite anchorsReview output

Coding workbench bridge

The build system and AI workbench are converging.

The adjacent coding workbench is the verification lane: inspect a codebase, plan changes, generate a patch, run tests, review diffs, and carry evidence back into SENTINEL. The result is one architecture for research, drafting, coding, and operational work.

Issue intakeRepo searchPatch proposalTest commandDiff reviewEvidence packet

Build proof

Visible progress without exposing private work.

Public proof should show capability, direction, and credibility — not private files, internal logs, or raw working notes.

SENTINEL image generation proof screenshot
Image generation in chat

What worked: generated asset displayed inside the conversation.

Why it matters: visual work can become governed, traceable, and reviewable.

What comes next: legal exhibits, diagrams, logos, and source-grounded visuals.

Governed chat interface

What worked: the product surface now shows conversational flow and asset display.

Why it matters: users need an approachable workbench, not an admin console.

What comes next: richer evidence rail, source cards, and approval cards.

Local files and RAG path

What worked: the architecture is shaped around approved folders, manifests, OCR, and anchors.

Why it matters: private work becomes searchable without surrendering control.

What comes next: document splitting, organization, and source-grounded answers.

Legal citation lane

What worked: legal work is framed around source status, citation status, quote spans, and review.

Why it matters: trust comes from verification, not confident prose.

What comes next: stronger official-source ingestion and filing-readiness blockers.

Where SENTINEL is going

A transparent roadmap without hype theater.

Progress is grouped by what is visible now, what is actively being connected, and where the larger architecture is headed.

Visible nowPublic brand site, governed chat preview, image proof, public ecosystem credits, Prefrontal architecture story.
In active developmentSource-grounded visuals, file organization, local document search, coding workbench bridge, legal review packets.
Planned expansionMultimodal research, enterprise governance, stronger local model lanes, official-source legal workflows, production trust center.

Open ecosystem

Built to orchestrate the best available tools responsibly.

SENTINEL is independent, but intentionally designed to integrate with open-source and public-interest software where licenses, security, and deployment policy allow.

Model and runtime lanes

Ollama, llama.cpp, vLLM, Text Embeddings Inference, Qdrant, ComfyUI, Diffusers, Axolotl, and candidate model families such as Qwen, Mistral, DeepSeek, BGE, E5/GTE, Nomic, Jina, and legal encoders.

View ecosystem

Legal authority plane

CourtListener, RECAP, Caselaw Access Project, eyecite, reporters-db, courts-db, Juriscraper, GovInfo, Congress.gov, eCFR, Federal Register, Open States, Maine official sources, and other source-specific connectors.

View legal lane

Security, policy, and evidence

OpenTelemetry, Open Policy Agent, OpenFGA, Keycloak, Presidio, Syft, Trivy, Gitleaks, Cosign, SOPS, Vault-style key management, gVisor, Firecracker, Promptfoo, Ragas, and observability systems.

View trust center

Trust center preview

Built for review, not blind faith.

The public trust center explains data handling, security posture, responsible disclosure, third-party acknowledgments, and status language in plain English.

TrustSecurityData handlingResponsible disclosureStatusCredits