Obsigen AI — How It Works

Pipeline

Five Steps to Outcomes

Understand

Parse the request and select the optimal strategy (Fast / ReAct / Swarm / Hybrid).

Plan

Break down into steps and choose tools — search, files, code, images, PPT.

Execute

Run actions in controlled environments — EEA-first routing, sandboxed code.

Verify

Critic/tester patterns or Swarm verification for accuracy and quality.

Deliver

Final output — report, SOP, deck, action plan — executive-ready.

Knowledge Layer

RAG — Retrieval-Augmented Generation

RAG allows Obsigen to answer using your knowledge, reducing hallucinations and improving traceability. Every answer can cite the source document, section, and version.

Recommended RAG Pipeline

Ingest: SOPs, policies, ERP docs, tickets, product catalogs
Process: Parse → chunk → metadata enrichment (dept, ACL, version, date)
Embed: Semantic vectors for fast similarity retrieval
Store: EU-hosted vector DB + metadata index
Retrieve: Top-k + filters (ACL, recency, system context)
Generate: Grounded response + evidence snippets

RAG Source	Example Content	Business Outcome
ERP / MRP	BOMs, routings, stock, lead times	Faster planning, fewer stockouts
SOC KB	Runbooks, detections, incident history	Faster triage and response
CRM / E-commerce	Product info, pricing, segmentation	Better next-best actions
SOP / ISO docs	Procedures, audits, compliance	Consistency + audit evidence

Customization

Trainable AI — Fine-Tuning

Fine-tuning adapts Obsigen to your terminology, tone, and process patterns. Teams get answers that match how your business actually works.

EU-Safe Training Posture

Train on approved datasets only
Keep training data & artifacts in EEA storage
Add evaluation gates and red-team prompts before deployment

Method	Best For	Notes
SFT / Instruction Tuning	Consistent "house style" answers	Needs curated examples
LoRA Adapters	Cost-effective specialization	Per-tenant or per-team variants
Preference Tuning	Align to internal standards	Uses ranked outputs
Tool-Use Tuning	Better planning/execution	Improves agent reliability

Infrastructure

Built on proven, sovereign foundations.

A pragmatic stack picked for EEA data residency, vendor independence and auditability — not the conference-talk cycle. Specific versions and CVE-relevant details are shared under NDA with serious buyers.

EEA inference

Open-weights LLMs running on European GPU infrastructure. NVIDIA-accelerated, with on-prem & sovereign-cloud options.

Containerised, portable

Docker & Kubernetes throughout. The same image runs on AWS, GCP, Azure, OVHcloud, Hetzner — or your own datacentre. No surprise lock-in.

Sandboxed execution

Tool calls run inside isolated, ephemeral sandboxes. Network egress allow-listed. Audit log on every action — written to your servers, not ours. We store nothing about the run.

Private web search

Self-hosted meta-search by default — no query telemetry leaves the EEA. Premium providers available on request, contracted under EU SCCs.

EEA data & vector store

Document storage in Warsaw, embeddings in a vector DB with row-level security. BYO-key (KMS / HSM) supported for regulated sectors.

Auditable — in your tenant

Every prompt, tool call, retrieval and policy decision is logged in your storage, under your keys, with your retention. We provide the audit primitives; you hold the data. EU AI Act Article 12 ready · SIEM-friendly (JSON / OTLP).

Exact versions, CVE-relevant components and tenant isolation specifics are shared under NDA — see Security.

How It Works