Deterministic AI Platform

AI subordinate to logic.

For decisions that cannot be wrong. When the generative model alone is not enough — and the result must be explainable, auditable and defensible.

01 — The Problem

Generative AI has become a risk in regulated environments.

Language models such as ChatGPT brought impressive fluency — and, with it, a class of problems that had no name yet in Law, Audit and Compliance: the confident wrong answer.

In regulated markets, this is not an inconvenience. It is exposure. Fines, disciplinary proceedings, loss of licence, annulled court rulings, leakage of protected data, irreversible dependence on a foreign vendor, invisible bias in critical decisions.

The market discovered, the hard way, that pure generative AI is not safe for decisions that must be defended.

May 2026 — TRT8 (Brazil)

BRL 84,000 fine

Lawyers fined for a filing containing case law invented by generative AI. A landmark case of prompt injection and hallucination in a court document.

May 2026 — STJ (Brazil)

Criminal inquiry opened

Justice Herman Benjamin ordered a criminal investigation into the use of generative AI producing fabricated content in court proceedings.

CNJ Resolution 615/2025

Reserved to human judgement

AI may not decide alone in a judicial act. Every decision requires documented human review.

Reliability Failure

Hallucination

The model invents a fact, a statute or a citation, with confidence. There is no native mechanism for the system to admit that it does not know.

Audit Failure

Opacity

The answer comes out with no trail. There is no way to reconstruct the path that led to that result, nor to audit what was used.

Sovereignty Failure

Data exposure

Big Tech APIs process sensitive data on servers outside the country. Data protection law, judicial secrecy and banking confidentiality all come into conflict.

02 — The Response

An architecture where the AI model never decides alone.

Kesshet has built a platform where generative AI has a clear and limited role: to read, to suggest and to flag for review. The decision belongs to an explicit, controlled and auditable logic layer.

The result is a system that delivers the speed of modern AI without the risk of modern AI. Every answer has a traceable origin in the source document. Every decision has a recorded reason. Every uncertainty is admitted rather than masked.

This approach has a technical name — neurosymbolic — but the name is not the point. The consequence is: defensible reliability, instead of seductive fluency.

03 — How It Works

Four principles that govern every Kesshet system.

Flow: document, AI reading, deterministic logic layer, auditable output
01

Offline by default

The system runs on your own infrastructure. Your data never leaves your house. No external API, no Big Tech, no dependence on a foreign server.

02

Auditable

Every decision has a traceable origin. Every source is cited. Every action is logged. The path from data to answer can be reconstructed at any time.

03

Admits what it does not know

When the system is not certain, it flags the uncertainty instead of inventing. An empty field is better than a wrong answer wearing the appearance of truth.

04

Learns from correction

Every adjustment made by a human expert becomes a rule of the system. The team’s knowledge is not lost — it accumulates in the machine.

04 — What We Solve

Applications built on the same architecture.

Each vertical below is an instance of the same platform, adapted to the client’s specific domain. The engine is single; what changes is the rule set and the type of document the system reads.

01

Tax Credit Recovery

Deterministic reading of statutory fiscal filings and tax documents to identify unclaimed credits. A result defensible before the tax authority, with a complete audit trail.

02

Compliance and Due Diligence

Analysis of high-volume PDFs — contracts, KYC, M&A files — with structured extraction, risk flagging and absolute respect for confidentiality. All processed offline.

03

Legal Risk Analysis

Reading of case files running to thousands of pages, identification of uncontested facts, pairing of the opposing versions of the dispute, and flagging of what requires human review. Defensible before the bar association.

04

Industrial Vision

Detection of non-compliance on CCTV cameras (protective equipment, operational behaviour, waste). Recognition by a proprietary model, with no image sent outside the plant.

05

Social Media Monitoring

Structured, deterministic analysis of comments and mentions on digital platforms. Consistent classification, without the noise typical of generative models.

06

Tax Reform Modelling

Impact simulation of Brazil’s tax reform (LC 214/2025, CBS/IBS) on real operations. Transition planning based on explicit rules — no model opinion, with verifiable calculation.

05 — Technical Proof

Engineering built and validated in real production.

Kesshet is not a slide deck. It is a set of systems that run, in cases with consequence. Below, some objective markers — without naming clients, out of respect for confidentiality agreements.

BRL 28.7M Tax credits identified in a single industrial operation, with a document-by-document audit trail.
F1 0.93–0.98 Accuracy metrics in industrial visual detection, validated in a plant environment with real cameras.
2.000+ Pages processed per legal analysis, with deterministic identification of contested and uncontested points.
100% offline The entire architecture runs on the client’s infrastructure. No sensitive data leaves the perimeter.

Kesshet also maintains a technical framework published publicly in an open-source software repository, available for audit and community use — concrete proof that the architecture exists and works.

06 — Who We Serve

Markets where error carries an irreversible cost.

Compliance and Financial

Banks, fintechs, asset managers. Where central bank regulation and data protection law require traceability of every automated decision. KYC and AML analysis, transaction monitoring and regulatory reporting — all defensible before the regulator, with a document-by-document audit trail.

Legal

Law firms and legal departments under pressure from the TRT8 and STJ precedents and CNJ Resolution 615/2025. Case analysis, documentary due diligence, structured filings — always with human review before signature, defensible before the bar association and the court.

Regulated Industry

Industrial operations under pressure from external audit, workplace safety and quality. Tax credit recovery, visual compliance monitoring, operational risk management — all on the client’s infrastructure, defensible before inspection and audit.

07 — Contact

For those who need more detail.

The technical presentation happens in conversation — not in a form. Write directly to the Kesshet team.

Team distributed between Tel Aviv and São Paulo. Specialists in AI architecture, digital law, regulatory compliance and industrial computer vision.