Talk to your Vista ERP, and trust the number that comes back
Viewpoint Vista is a 20-year-old construction ERP only a DBA can safely query, and a generic LLM guesses its schema, producing confidently-wrong reports with no error message. VistaGPT encodes two decades of Vista tribal knowledge into a verified knowledge engine and deterministic validators, then turns a plain-English question into validated SQL, an automated Excel workbook, or a published Power BI dashboard. Finance teams stop waiting on the DBA, and the consultants who get paid to build those reports, workflows and dashboards do it in a fraction of the time. It is engineered to refuse rather than guess.
VistaGPT in development, replayed end to end: ask in plain English, watch the workflow run and check off, and read back validated T-SQL grounded in the real Vista schema.
Vista's schema is famously treacherous: cryptic 4-letter table codes (APTH, JCJM), composite join keys, case-sensitive collation, mandatory WITH (NOLOCK) hints, and soft-delete columns that silently drop rows. So finance and operations wait days for a DBA to write a report, and institutional knowledge lives in brittle Excel and Crystal Reports built by people who have since left.
Pointing a generic LLM at it is worse than useless: it guesses at Vista's schema, returning confidently wrong SQL with no error to warn you. On financial data, that's not a glitch; it's a liability.
The same plain-English question can come back as validated SQL, a rebuilt Excel workbook, or a live dashboard.
A validated SELECT, every table/join/filter justified and confidence-scored, copy it and run.
Workbooks treated as code: VBA, Power Query and formulas extracted, AI-refactored, validated and injected back.
A prompt, a SQL file, or a screenshot, becomes a published, refreshing Fabric dashboard.
The same platform pays off twice, for the firms that run on Vista, and for the consultants who get paid to make Vista work for everyone else.
Report builds, workflow changes, dashboard delivery, workbook modernization, the high-value Vista work clients pay a premium for, and that normally only a senior consultant can do. VistaGPT compresses it: deliver senior-grade output from junior staff, modernize whole client portfolios at once, and productize hours of expert time into minutes. Encode your best consultant once, and run it across every account.
Controllers, accountants and project managers ask the ERP in plain English and get a trusted, validated answer back, SQL, an Excel workbook, or a dashboard, without filing a ticket or waiting days. Self-service analytics your finance team can actually rely on, because the system refuses to hand back a number it can't prove.
Every query is forced through a deterministic, zero-LLM validation gauntlet before a human sees it. If it can't be verified against the real schema, it isn't returned.
The engine is set in configuration, with AgentSDK as the default. It is not picked automatically per query, and users make no engine decision. Validation so far is internal only, and no accuracy score is published.
VistaGPT is not one model behind a chat box. It is eight services across two repository generations, four SQL engines, a schema-intelligence layer, two augmentation surfaces, and a gateway, all reading one centralized knowledge base. Every cell below is a service in the Codelab-era codebase.
The workhorse for multi-table domain reasoning. CrewAI Flow orchestration; a DomainKnowledgeCrew with recall_joins and recall_objects tools that retrieve verified join patterns from the KB instead of inferring them. Mem0 multi-turn memory; guardrails reject any object outside the manifest.
The default engine: a 13-stage hybrid pipeline. Stages 2-7 are six parallel deterministic ChromaDB lookups with zero LLM involvement; Stage 9 Reconcile drops hallucinated schema before the Generator writes SQL. sqlglot validation with up to two auto-retries.
A 10-stage quality pipeline grounded in a reference corpus of real production SQL, not synthetic. Stage 4 finds the closest Crystal Report match before generation; a stage0→stage10 smoke-test gate covers every stage. An async stage-10 learning pipeline caches semantically after each query.
The schema-intelligence layer every engine depends on. A 19,249-entry manifest covering every table, view, procedure and column in Vista. When AgentSDK's Reconcile asks "does this column exist?", it asks VGPT3. 13+ MCP tools; a Gen-2 REST API; no external API keys.
AI inside the Excel teams already use. Five workflows, expert analysis, workbook generation, AI refactoring through 6 hard validation gates, batch processing, and SQL-from-Excel-exports. The engine is set in configuration, with AgentSDK as the default.
A FastAPI MCP server exposing seven Power BI workflows to the pipeline. WF10, the PBIR builder, assembles a full report project, data model, DAX measures, layout, from AI-generated SQL via a --sql flag. Design target: a report build that takes about 12 hours hand-held today, brought down to minutes.
A FastAPI gateway routing all eight services through one engine registry, plus a React 18 / TypeScript front end. May 2026 added multi-route SSE (parallel streaming from several engines at once), in-chat file uploads, and PBI proxy routes.
One centralized platform_kb submodule, synchronized across all eight services. Each service makes HTTP calls to the KB instead of running a local ChromaDB. Add a join pattern once and it is live for CrewGPT's recall_joins, AgentSDK's evidence stages, and VistaForge's corpus at once.
VistaGPT is the construction build of our AI data-intelligence work, the same grounding-and-refusal discipline we apply to any system that has to be right about numbers.
A generic chatbot would have to re-derive, and get wrong, twenty years of tribal knowledge. This is that knowledge, made into a system.
CrewGPT, AgentSDK, VGPT3 and the next-gen VistaForge sit behind one JSON registry, add or swap an engine with a single edit. The engine is set in configuration, with AgentSDK as the default; it is not picked automatically per query.
Refuse rather than guess. Every generated query is forced through ~19 deterministic, zero-LLM Vista rules, NOLOCK, composite-key joins, fan-out, soft-delete/purge, posting-path NULLs, and verified against all 70,641 real columns before a human ever sees it.
Before writing new SQL, the engines reuse 1,200+ proven production Crystal Report queries, so an answer can cite a real, already-trusted report, not an invention.
9 agents across 5 crews with a SchemaRetriever grounding over 70,641 columns, a HallucinationGate (5 checks + auto-fix), and self-improving memory that compounds accuracy with use.
51 tools, 5 workflows: extract → analyze → AI-refactor → validate (official Microsoft Power Query parser) → transactional inject with rollback → COM-verify.
A single Vista knowledge hub, 56,080 files, 809 MB of object docs, schema, reports and traps, serves every engine over one interface. Edit once, propagate everywhere; no answer drift between engines; ~1.5 GB of duplicated copies eliminated.
Six structural guarantees stand between a question and an answer. None is a prompt or a hope, each is a deterministic gate or a grounded source of truth that runs before a number reaches a human.
Every engine reasons over the VGPT3 manifest, every table, view, procedure and column in Vista. An object that is not in the manifest cannot appear in generated SQL, so the system cannot fabricate a table or column name.
VGPT3 · schema lookups · 743 tests passing
In the AgentSDK path, Stages 2-7 run six parallel ChromaDB lookups, tables, approach, dates, columns, joins, filters, with no LLM writing anything. The evidence boxes contain only patterns that already exist in the knowledge base.
AgentSDK · 6 parallel evidence stages · no LLM
Stage 9 is a deterministic Python process: it compares every column, join and filter in the proposed blueprint against the platform KB and drops anything not verifiably present, before the Generator LLM writes a single line of SQL.
Stage 9 Reconcile · the second deterministic firewall
The final SQL passes through a sqlglot validator with up to two automatic retries for syntactic correctness. A query that cannot be parsed and verified is not what gets streamed back to the user.
sqlglot · ≤2 auto-retries · syntactic gate
All eight services read a single centralized platform_kb over HTTP rather than per-service ChromaDB copies. Edit a join recipe once and every engine sees it. There is no version of the schema that one engine trusts and another does not.
platform_kb · 56,080 files · single source of truth
Before inventing SQL, VistaForge reuses 1,200+ production Crystal Report queries so an answer can cite a report that is already trusted. The pipeline's stage0→stage10 smoke tests and VGPT3's 743 tests keep that grounding honest as the system changes.
1,200+ Crystal Report corpus · stage0→stage10 gate
VistaGPT is in development with internal validation only. Below are design targets and sourced facts, not client results.
Design target, not a client result: a Power BI report build that takes about 12 hours hand-held today, brought down to minutes.
A platform that refuses to guess earns that posture in commits and tests, not in slideware. Here is the engineering record so far.
The Codelab era alone covers a V3 CrewGPT rewrite (407 commits), a 13-stage AgentSDK pipeline built from scratch, VistaForge's 10-stage quality pipeline, a full front-end and back-end rebuild, and a new Power BI MCP service, inside a two-month window.
On May 22, 2026, LiteLLM proxy and Langfuse tracing were merged across all eight services at once. Every LLM call in every pipeline stage is now traced end-to-end, token spend, per-stage latency and model routing are observable at the platform level, not per service.
Validation leans on Microsoft's own parsers and schemas as the authority, making "wrong-but-plausible" output structurally hard to ship.
Whether your team waits days for reports or you're a consultancy that builds them for a living, we'll build a trustworthy conversational layer over the database, grounded, validated, and engineered to refuse rather than guess.