Project 01 · Agentic AI · Data Intelligence · Application Generation · AI Engineering
Turn organizational data into context, and context into software.
Organizations already have enormous amounts of useful data, but it is fragmented across files, databases, cloud storage, documents, reports and internal systems. SKatalyst AI is an attempt to build the layer between that raw organizational data and the software people actually need.
Instead of beginning with an application and manually wiring data into it, SKatalyst begins with the organization’s existing information. It ingests it. Understands it. Structures it. Builds context around it. Then uses that structured understanding to generate useful applications.
01 · The problem
Most organizations already hold the information they need. It just exists as separate files and rows, spread across drives, mailboxes, databases and tools that were never designed to explain each other.
A person understands the relationships implicitly: this spreadsheet belongs to that project, this report describes that test, that table holds the measurements, this document defines the process. Software normally sees none of it. SKatalyst attempts to reconstruct that missing organizational context, and to keep it.
02 · The idea
The whole company in one line: the value is not the model. It is the persistent, validated understanding that sits between an organization’s data and the software built on it.
Thousands of disconnected objects
SKatalyst AI
persistent context · human approval · evidence
Working software
Bring organizational data in from where it already lives: uploads, drives, buckets and databases.
Scan read-only. Extract text and tables, OCR images, read schemas, metadata and relationships.
Propose a structure: entities, a schema, a folder taxonomy, quality findings and a recommended architecture.
Persist the result in PostgreSQL so the understanding outlives any single AI conversation.
Discover KPIs the data can actually compute and decide which modules expose them.
Build working application components: dashboard, CRUD, report, RAG chat, plus a database and API.
Compile a test oracle from the data model and compare generated behaviour against reference values.
Preview in an ephemeral runtime, export a Docker package the customer owns, verify, roll back if needed.
03 · How it works
SKatalyst is not built to upload files to a language model and ask questions. Each run turns input data, progressively and inspectably, into persistent structured context.
A read-only scan produces an inventory the user approves. A structuring plan proposes entities, a schema and a folder taxonomy. An architecture recommendation explains its confidence and what was excluded from the estimate. KPI discovery only offers metrics the data can compute. Every step is stored in PostgreSQL with lineage back to the source files, and the context a run used is frozen and hashed at start so it can be reconstructed later.
The LLM should not be the database.
04 · The intelligence layer
SKatalyst combines deterministic software with AI reasoning, and keeps the boundary visible. The default platform provider is a deterministic heuristic; a model is added when a key is configured, and features that need one degrade to a non-AI path instead of failing.
Guaranteed by code, tests and the database.
Where interpretation is genuinely required.
Software guarantees what can be guaranteed. AI handles what requires interpretation. Neither is assumed correct: proposals carry confidence, unavailable metrics are marked unavailable, and a human confirms before anything is built.
05 · System architecture
Verified against the platform’s source. Each layer only depends on the ones above it; validation sits between generation and delivery on purpose.
12 source types. Cloud and database connectors read metadata today; execution stays Preview until proven against a real account.
Every run starts with an inventory the user approves.
The understanding is stored, versioned and traceable, not re-derived per prompt.
Every AI step has a non-AI path or a BYOK gate; the platform never silently guesses.
4 generators shipping; 10 module types declared.
The oracle never reads generated code, so a generator error cannot hide itself.
Export is blocked while validation is failed. That block is intentional.
Connectors
SKatalyst doesn’t require the organization to move everything into a new ecosystem before it can begin understanding the data.
Twelve source types are implemented, from local upload to Databricks. The labels below are the product’s own: a connector may not leave Preview until a real byte-read against a real external account has been proven and recorded. Connecting and browsing already work; local upload runs the full pipeline today.
06 · What it can build
Structured operational interfaces over the detected entities, with generated API endpoints for each.
ShippingKPI cards and charts built only from metrics the data can compute; unavailable metrics are marked, never faked.
ShippingNatural-language access grounded in the organizational context. Runtime keys stay with the customer.
ShippingStructured reporting from the same context, exported with the application package.
ShippingData-oriented analytical modules where the underlying data supports them.
RoadmapDeclared module types waiting for generators.
Roadmap
07 · Trust and validation
If AI is going to generate software from organizational data, generation cannot be the final step. SKatalyst treats validation as part of the architecture: expectations are compiled from the data understanding, never from the generated code, so a generator’s mistake cannot vouch for itself.
TestOracleCompilerTurns the confirmed KPIs, entities, schema and business context into behavioural assertions. Pure and deterministic: no database, no network, no model, and never a line of generated code.
ReferenceExecutorEvaluates each assertion against the ingestion-time sample snapshot: sums, counts, averages, minimums and maximums. Read-only; never opens a live connector.
BehavioralProbeRunnerBoots the app in an isolated runtime seeded with the same sample and probes its endpoints, including negative tests that must be rejected.
ComparatorExact or relative comparison of observed versus reference values, so a rounding difference is not a failure and a wrong KPI is.
ValidationReportServiceEvery assertion carries provenance: which KPI, which table, which formula. The report is the evidence a human approves against.
ValidationOrchestratorRuns compile, reference, probe and persist; falls back to a reference-only report when the runtime is unavailable; always tears the runtime down.
Human approval
The system proposes; a person decides at the moments that matter. Two gates are distinct by design: approving a specification before anything is built, and accepting a preview after it is built.
Solution lifecycle · 12 states
Highlighted states are the human gates. Transitions are an explicit table; illegal ones are refused, repeated ones are idempotent, and unarchiving never restarts a runtime.
08 · Product experience
The in-app stepper tracks eight steps on every run page. Each step explains what the system knows, why it matters and the next best action, in the product’s own voice.
09 · Engineering
Verified from the repository, organized by purpose. The platform is a Node and PostgreSQL system with a small, explicit AI surface.
SKatalyst is not tied to one model provider. Tenants bring their own keys for Claude, OpenAI or Gemini; keys are encrypted at rest, decrypted only at execution, and one active key per provider is enforced. Generated chat applications use the customer’s own runtime provider, chosen at deploy time, never requested during design.
The value sits around the models: the context, the workflows, the structured data, the application architecture and the validation system.
All figures counted from the repository on 28 August 2026: 180 commits since June 2025. Endpoint count is a floor (391 in the main router plus 11 route modules).
10 · Evolution
Initial project structure.
Document ingestion pipeline, LLM-assisted classification with a heuristic fallback and a vendor-neutral provider registry.
Metrics model, blueprint generation and the first artifact export: from organizing data to generating software within days.
Multi-tenant schema, OAuth, encrypted vault and BYOK, OCR and visual understanding, ten connector backends, immutable context snapshots, orchestrated runs and KPI discovery.
A self-critical audit listed the gaps: beta gate, rate limiting, headers, validation library, unproven isolation. Most were closed in the following weeks.
The agentic validation loop was disclosed and built in increments: oracle compiler, reference execution with anomaly detection, behavioural probes. Patent package filed.
The Solution entity with its 12-state machine, two distinct human gates and transactional audit events.
Landing admin, rate-limit counters, deploy and verify runbooks, onboarding, and the frozen authenticated app for the September 2026 beta.