PLANCK AI

Planck AI

A custom LLM that gets smarter with every transaction.

Not a chatbot wrapper. A ReAct agentic system with 9 tools, dual-mode intelligence (Chat + SQL), multi-provider failover, and a closed-loop fine-tuning pipeline. Ask in English. Get answers from your actual data. Watch it learn.

Built-in tools

Agent loop depth

AI providers

$3/mo

Self-hosted at scale

What it does

Chat Mode — conversational operations

Create contacts, schedule tasks, search deals, aggregate pipeline data — all through conversation. "Add John from Acme as a lead and schedule a follow-up for Friday." Done in one message.

Intelligence Mode — text-to-SQL analytics

"What was our close rate by region last quarter?" → SQL generated, validated, executed, results explained in English. Schema-aware: queries scoped to your actual tables and tenant.

ReAct agentic loop

Reasoning + Acting: the AI generates a response, decides if it needs tools, executes them, reads results, and continues reasoning — up to 5 iterations per request. Not scripted. Truly autonomous.

SQL validation & tenant scoping

Every generated query is validated: SELECT-only, no injection, column existence checked, tenant scoping enforced. Multi-tenant by design — impossible to access another company's data.

SSE streaming with tool visibility

Real-time streaming via Server-Sent Events. You see text appear token-by-token, tool calls fire, SQL execute, and results flow back. Full transparency into the AI's reasoning.

Multi-provider with automatic failover

Qwen3 (self-hosted), Claude (Anthropic), GPT (OpenAI), K2.5 (Moonshot) — all unified. If one goes down, traffic routes to the next in <100ms. Zero interruption.

Feedback-driven fine-tuning

Thumbs up/down on every response. This feeds the Tinker pipeline: Langfuse traces → training data → fine-tuned model. Your AI gets measurably better every week.

Sensitive data redaction

Emails, passwords, API tokens — automatically redacted from AI responses. Your data is queryable but never exposed in its raw form through the chat interface.

How it works

01

You type a question in plain English

"Show me all overdue invoices from Q4 and draft follow-up emails for each." — Chat mode or Intelligence mode selected automatically based on intent.

02

Model router selects the best provider

Based on task complexity, cost budget, and provider health. Self-hosted Qwen3 for routine queries, Claude for complex reasoning, GPT for code generation.

03

Agentic loop reasons and acts

The AI thinks, decides it needs data, calls query_database with generated SQL, reads 47 overdue invoices, drafts follow-up emails, calls create_task for each.

04

Results stream back in real-time

You see every step: tool calls firing, SQL executing, results flowing, text generating. Full transparency. No black box.

05

Feedback closes the loop

Rate the response. Your feedback + the full trace → Tinker fine-tuning pipeline. The same question asked next month gets a better answer.

Why it's different

ReAct agentic loop with 9 real tools — not a prompt-and-pray chatbot

Schema-aware SQL generation scoped to your actual database structure

Self-hosted Qwen3 option: zero API costs, zero data leaving your network

Fine-tuning pipeline turns every interaction into training data for better models

Multi-tenant SQL validation makes cross-company data access impossible by design

SSE streaming shows tool calls in real-time — full transparency into AI reasoning

Ready to try Planck AI?

Start Building →