Cloud Native Architecture

Built for Enterprise Scale

Agomonia Labs is designed with a modular, cloud-native architecture that separates stateless request handling, durable memory, retrieval indexes, file storage, agent execution, evaluation, and governance, allowing every layer to scale, evolve, and operate independently.

Architecture in motion Scale, retrieval, governance

01 / Cloud Entry

Firebase Hosting delivers the product experience close to users.

React and Vite frontends are served through Firebase Hosting, while authenticated requests move through HTTPS and JWTs into backend services that can evolve independently.

React UIFirebase HostingJWT

02 / Runtime

Cloud Run scales stateless APIs and agent execution.

FastAPI services run on Cloud Run, so request capacity can scale horizontally while sessions, documents, vectors, memory, and logs remain in durable managed stores.

FastAPICloud RunAsync agents

03 / Retrieval

Hybrid retrieval combines semantic and exact-match intelligence.

Cloud SQL, pgvector, Firestore, full-text search, chunk storage, and source files work together so grounded answers can cite private documents and operational data.

pgvectorFirestoreHybrid search

04 / Video Pipeline

Large video files use a production-aware ingestion path.

DocIntel routes larger videos through direct browser-to-GCS upload, then processes metadata, sampled frames, transcript signals, captions, timestamped segments, and embeddings without treating video as a disconnected media silo.

Direct GCS uploadffmpeg / ffprobeTimestamp chunks

05 / Agent Layer

ADAR orchestrates specialist agents and domain workflows.

Google ADK, Gemini, domain tools, workflow agents, and memory services coordinate work across healthcare, restaurants, sports, music, real estate, finance, video intelligence, and document intelligence.

ADARGoogle ADKSpecialist agents

06 / Governance

Evaluation, traceability, and approval make AI reviewable.

MFA, trace IDs, audit logs, response scores, feedback, usage controls, privacy flags, RBAC, and approval gates help teams operate AI workflows with confidence at scale.

MFAAudit logsRBAC

DocIntel Platform Architecture

From unstructured documents and videos to cited answers and workflows.

DocIntel is Agomonia Labs' intelligence platform for turning unstructured documents, transcripts, audio, and videos into reliable business answers. It combines OCR, layout understanding, table extraction, video metadata processing, frame sampling, timeline chunking, semantic retrieval, and LLM reasoning to support question answering, summarization, comparison, video Q&A, and domain-specific workflows. Every response is grounded in retrieved evidence and can be traced back to the original source, including timestamps for processed videos.

01Documents + Videos

PDFs, scans, forms, reports, contracts, policies, recordings, trainings, meetings, walkthroughs, and enterprise media.

02Extract + Structure

Text, tables, layout, sections, entities, dates, video metadata, frames, captions, transcripts, and timestamps are understood.

03Knowledge Index

Extracted knowledge is chunked, embedded, and stored for hybrid retrieval.

04AI Reasoning

AI retrieves the right evidence for answers, summaries, comparisons, and structured outputs.

05Cited Answers + Workflows

Outputs stay traceable and can trigger repeatable business processes.

Document Intelligence Lifecycle

Simple experience. Powerful engineering underneath.

DocIntel presents a simple experience for business users while powerful engineering works underneath. From the outside, the flow is easy to understand: enterprise documents and videos are ingested, important content is extracted and structured, knowledge is indexed for search, and AI uses that trusted context to generate cited answers, timestamped video answers, summaries, comparisons, and workflow outputs.

Behind that simple experience is a deeper technical foundation: direct-to-GCS large video upload, ffprobe metadata extraction, bounded ffmpeg frame sampling, optional visual captioning, optional audio transcription, OCR, layout parsing, table extraction, deterministic rules, regular expressions, semantic chunking, vector embeddings, metadata storage, hybrid retrieval, agent orchestration, evaluation, audit logs, and governance controls. This balance lets DocIntel use rules for precision, embeddings for discovery, AI for reasoning, workflow orchestration for consistency, and human review for trust. The public website keeps the diagram business-friendly, while detailed architecture views can be used in technical documentation, engineering discussions, investor material, or enterprise implementation reviews.

1Ingest

Upload PDFs, scans, forms, reports, contracts, policies, enterprise documents, transcripts, audio files, and videos.

2Understand

Extract text, tables, layout, sections, entities, dates, video metadata, sampled frames, transcript signals, captions, and timestamps.

3Index

Chunk, embed, and store knowledge in a searchable retrieval layer.

4Reason

Retrieve evidence and generate answers, summaries, comparisons, and structured outputs.

5Verify

Trace answers back to documents, pages, sections, and extracted evidence.

6Automate

Turn document intelligence into repeatable domain workflows.

Summarization

Condense long documents into executive summaries, section summaries, risks, obligations, timelines, and key decisions.

Comparison

Compare contracts, policies, reports, claims, proposals, or versions side by side with highlighted differences and cited evidence.

Video Intelligence

Convert recordings into metadata, frame samples, captions, transcripts, timeline chunks, embeddings, timestamp-aware citations, and video Q&A.

Domain Workflows

Apply business-specific rules, checklists, and output schemas for legal, insurance, finance, healthcare, operations, and compliance use cases.

Identity Management and Security

Secure signup, login, MFA, and role-aware access for DocIntel.

DocIntel identity management is designed so users can enter the platform through a controlled signup and login flow before they reach sensitive document intelligence, chat, workflow, billing, or administrative areas. Signup captures the account information needed to create the user profile and connect the user to the right workspace, while login verifies the user before any protected DocIntel capability is available.

Multi-factor authentication adds a second verification step beyond the password. This reduces the risk of unauthorized access and makes DocIntel more appropriate for document-heavy and workflow-heavy environments where teams may handle contracts, healthcare documents, lease records, restaurant operations, billing context, or other sensitive business information.

After authentication, DocIntel uses workspace scoping, role-based access control, audit history, and governance-aware workflow rules so each user sees only the data and actions allowed for their role. Admins, contributors, reviewers, operators, clinicians, patients, restaurant owners, staff, and other personas can be mapped into permission boundaries while traceability remains available for security review.

Open identity demo
01Signup

Create a user account and connect the user to the appropriate workspace or product access path.

02Login

Validate user credentials before protected document, chat, workflow, billing, and admin capabilities are available.

03MFA verification

Require an additional verification step to reduce unauthorized access risk.

04Role-aware access

Apply RBAC, workspace scoping, audit history, and governance rules after authentication.

ExperienceReact / Vite UIFirebase Hosting
HTTPSJWT
API RuntimeFastAPI BackendCloud Run
Asyncrun_async
Reasoning CoreADAR OrchestratorGoogle ADK + Gemini
Agent LayerSpecialist Agentsdomain tools and workflows
Retrieval StoreCloud SQL + pgvectorsessions, chunks, vectors
Workspace MemoryFirestorememory, tenants, logs, indexes
Source ArtifactsGoogle Cloud Storagesource files, videos, frames, and extracted chunks
Video Processingffmpeg + ffprobemetadata, frames, captions, timelines
GovernanceEvaluation Layerjudge scores, traces, feedback

Stateless API scale-out

FastAPI runs on Cloud Run, so request capacity can scale horizontally while sessions, documents, vectors, and logs stay in managed stores.

Tenant and workspace isolation

Users, workspaces, documents, chat history, vectors, and agent runs are scoped so teams collaborate without crossing data boundaries.

Async agent runtime

ADK runners execute asynchronously, letting specialist agents call tools, retrieval systems, and live services without blocking the app.

Indexed retrieval layer

pgvector, Firestore vector indexes, and full-text search keep semantic and exact-match retrieval fast as knowledge volume grows.

Video-aware knowledge base

Video timelines are chunked, embedded, and cited with timestamps, so users can ask what happened in a recording and connect the answer with the rest of the workspace knowledge.

Streaming user experience

Long-running chat, summaries, and workflow outputs stream progress to the browser while backend services continue work safely.

Operational governance

MFA, trace IDs, audit logs, response evaluation, feedback, usage limits, privacy controls, RBAC, and approval gates keep scaled AI reviewable.