Knowledge that isn't shared doesn't compound.
We built BiVelio by running real companies. Every algorithm we design comes from a concrete problem, not a paper. We believe explaining it in the open is the most honest way to make progress.
This section is our research notebook: here we publish the ideas, models and technical decisions behind the product. We want anyone — customer, engineer or curious mind — to understand how the system they trust to automate their operation actually reasons.
In the open
We publish the reasoning, not just the result. Transparency is a requirement, not a marketing gesture.
With rigor
Math, code and real measurements. If we claim something, we show how we know.
That compounds
Each article builds on the last. Knowledge that accumulates is what creates long-term advantage.
To share
We write to teach. If an idea helps another team build better, it has done its job.
Research articles
Browse by topic. Each piece is self-contained: context, intuition, formalization and results.
- Algorithms
BV-OCR: inside the document-reading engine
BV-OCR is BiVelio's document-reading engine: the in-house stack that feeds the Documents module. It is three coordinated pieces. A deterministic GPU document parser, written in C++ on CUDA and TensorRT, turns every page into verbatim text with coordinates, layout, tables and formulas at hundreds of images per second. A compact vision-language model proposes field-level interpretations on top of that text: which value is the invoice total, which date is the due date. And the value-in-OCR gate requires every proposed value to be anchored to a span of the OCR-recognized text through an evidence locator; anything that cannot be anchored is rejected and routed to human review. This article walks through the three pieces, presents the numbers from our reproducible benchmark harness over public datasets, and explains why this split turns model output into evidenced data.
12 min read - Systems
From documents to governed data: inside the Documents module
Most of a company's operational data is born in documents —invoices, contracts, forms, scanned mail— and dies there, unstructured. BiVelio's Documents module turns those documents into validated, evidenced, actionable data inside the customer's governed operation: ingestion and OCR on BiVelio-controlled infrastructure, field-level extraction where every value carries its provenance and a confidence signal, a human review console for anything doubtful, and governed automation with permissions, an audit trail and human gates on critical steps. This article walks through that pipeline and makes its central case: evidence beats bare confidence scores, because a person can verify in seconds what a percentage can only promise.
8 min read - Agents
RPA vs workflow automation vs AI agents: layers, not rivals
RPA, workflow automation and AI agents don't compete: they are distinct layers of the same problem. RPA mimics clicks for stable, high-volume tasks; workflow automation orchestrates predefined sequences across systems; AI agents plan and reason over ambiguous, unstructured inputs. Each paradigm solves one stretch of the determinism-to-autonomy spectrum, and real operations need all three. What was missing wasn't more automation, but a layer that governs what runs on its own and what a human decides.
11 min read - Systems
A single agent fails: governed AI in 3 layers
Most enterprise AI bets fail because they put a single agent in charge of everything: remembering, diagnosing, deciding and executing. BiVelio separates those responsibilities into three composable layers under governance: a Brain that remembers, Workers that diagnose and Agents that execute — with a Trust Layer that governs every consequential action. This article explains the three-layer architecture, how it composes end to end, how it differs from an isolated LLM agent, from RPA and from workflow automation, and how the Autonomy Console measures and steers the Autonomy Rate.
9 min read - Systems
AI agents: it's not the model, it's the architecture
Enterprise AI agents do not fail because of a weak model, but because the operational architecture around them lacks memory with source traceability, authority limits and an audit trail. This article defines what a governed enterprise agent is, offers a taxonomy of risks in five families, the controls that turn each risk into governed autonomy and a five-layer reference architecture. The thesis: it is not autonomy that decides the outcome, it is architecture.
9 min read - Foundations
Your ungoverned AI agent is a liability, not a milestone
Automation is not a switch: it is a spectrum that runs from the scheduled, deterministic task to governed autonomy, where AI executes the repeatable and people decide the critical. This article walks through the four stages of maturity —scripted automation, BPM orchestration, AI agents and governed autonomy—, explains why agency without governance is a liability rather than a milestone, and shows how BiVelio operationalizes governed autonomy as a layer on top of the tools the company already uses.
11 min read - Agents
How to govern AI agents: authority, not output
Governing an AI agent is not about reviewing what it produces, but about governing its authority: who it is, what it can execute on its own, what it must escalate to a person, and how an error is reversed. Six primitives —roles, authority thresholds, policies, approvals, audit, and rollback— turn autonomy into governed autonomy. This article defines those primitives, compares the governance models, and explains how BiVelio applies them as a governed autonomous operations layer that connects on top of the tools a company already uses.
10 min read - Foundations
Human-in-the-loop: 5 components, not an approval button
Human-in-the-loop is not an approval button bolted onto a model: it is a complete operating model. It is a Trust Layer of permissions, authority thresholds, approvals, full audit and rollback that governs how autonomous work is executed. Now that agents act —not just suggest— the difference between a feature and an operating model decides whether autonomy is governed or uncontrolled. This article defines that model, contrasts it with oversight-as-a-feature, describes its five components, and explains how BiVelio materializes it in its Trust Layer inside a governed autonomous operations layer.
11 min read - Systems
Buying more software won't make you an autonomous company
Between what a company knows and what it executes autonomously there is a layer that does not exist in the modern stack today. Email moves messages, the CRM stores contacts, the ERP records transactions, RPA copies data and agents generate text — but none of them turns a company's knowledge into governed autonomous operations, with memory, decisions and accountability. That is the missing layer. We explain what it must do, why it is empty today and how BiVelio builds it as a layer that connects on top of your tools rather than replacing them.
11 min read - Foundations
Adding a copilot is not an operating model
A human + agent operating model is an organizational design in which people and governed AI agents share the same work, the same decision rights and the same audit trail — and where autonomy is set deliberately per process, not left for tools to decide on their own. It is not a question of which software to buy, but of how processes, responsibilities and control are redesigned when agents join the workforce. This article defines that model, compares it with the classic company and with automation, and explains how BiVelio materializes it: Brain, Workers, Agents + Velio, Trust Layer and Autonomy Console.
11 min read - Foundations
RPA automates tasks. A GPOS governs operations
A Governed Process Operating System (GPOS) is an operating model, not a software suite: it turns a company's knowledge into governed autonomous operations, where AI executes the repeatable and people decide the critical. This article defines the category, contrasts it with RPA, BPM, AI agents and iPaaS, and explains how BiVelio implements it as a governed autonomous operations layer built on a Brain, Workers, Velio and agents, a human-in-the-loop trust layer, and an Autonomy Console.
10 min read - Foundations
AI Workforce OS: not a suite, but an operating model
An AI Workforce OS is an operating model in which a governed digital workforce turns a company's knowledge into autonomous operations — not an all-in-one software package that replaces your tools. We explain the category, the problem it solves and how BiVelio delivers it as a governed autonomous operations layer built on five pillars: Brain, Workers, Agents + Velio, a human-in-the-loop Trust Layer and an Autonomy Rate measured in a single console.
10 min read - Agents
95% of AI pilots fail: it's not the model
AI agents rarely fail because the model is weak: they fail because they lack operational context, governance and human control points in the real operation. We walk through the five failure modes (context, coordination, governance, verification and value), why they are structural and not a prompt problem, and why the durable solution is an operating model —living memory, governed agents and humans in the loop— and not a smarter prompt.
12 min read - Systems
40% of agentic AI dies without an Autonomy Rate
The Autonomy Rate is the share of a company's operational work that runs autonomously and within policy: the single KPI of governed autonomy. Unlike automation metrics, it only counts work bounded by permissions, authority thresholds, approvals, audit and rollback. We explain what it measures, how to define it, how it differs from automation coverage or DORA metrics, and how BiVelio turns it into the headline metric of its Autonomy Console.
12 min read - Retrieval
Your RAG's problem isn't the model, it's the order
Vector search is great at fetching candidates and mediocre at ordering them. That's why serious retrieval has two stages: one fast and approximate, one precise and expensive. We cover why reranking more than doubles ranking quality (with data), the hidden cost almost nobody counts, how the landscape stacks up (Cohere, Voyage, Pinecone, ELSER, ColBERT, Jina), and how we build a reranking index that lives only as long as a query — and why that makes us better.
8 min read