Technology
How Kaleido Engine works.
Every part below carries a status: Live, In build or Planned. We would rather show the roadmap than let it read as finished.
Architecture
From request to quote.
- 01
Request
Voice, text, file or photo
- 02
Agents
Parse, standardise, classify
- 03
Pricing engine
Deterministic line prices
- 04
Human check
A specialist verifies every line
- 05
Quote
Formal quotation to the buyer
What is live, and what is next.
RFQ intake
Turns voice, text, files and photos into structured requirements.
LivePricing
A deterministic engine calculates line prices. AI-sourced estimates are marked as indicative.
LiveModels today
Google Gemini Flash models, through Google’s Gemini API.
LiveRetrieval
Word matching across the price list, active suppliers and past quotations. Embeddings and vector search are planned.
LiveEvaluation
Every release is tested on a held-back set of 10 requirements (20 lines). Latest, 3 November 2026: 100% of lines found, quantities right on 100% and categories on 100% of those. Known limit: the set is small, and one case (a welcome kit split into its contents) led to a rule fix before this run, so it is no longer unseen.
LiveSelf-hosted inference
Open models served on dedicated GPU instances using NVIDIA NIM, TensorRT-LLM and Dynamo Triton, so client data stays inside Kaleido’s environment.
In buildPhoto-to-BOM
Multimodal models identify components and materials from a photo or drawing.
In buildDomain models
Fine-tuned models for HS code classification, starting with NVIDIA NeMo.
Planned
Roadmap
The agents behind every requirement.
Phase 1 runs today on our own code and Google Gemini. The NVIDIA column is where each agent moves once self-hosted inference is live. The pricing engine stays deterministic, and a person still approves every quote.
| Agent | What it does | Runs on today | NVIDIA target | Phase | Status |
|---|---|---|---|---|---|
| Orchestrator | Runs every requirement through the same stages and hands off to a person with a list of what to check | Kaleido code | NeMo Agent Toolkit | 1 | Live |
| Document intake | Reads RFPs, BOQs, scans and tables | Gemini, spreadsheet and Word parsers | Nemotron Parse | 1 | Live |
| Specification | Standardises item text into structured specs | Gemini Flash | Open model on NIM, TensorRT-LLM | 1 | Live |
| Retrieval | Finds stocked items, suppliers and past prices | Word matching on our database | NeMo Retriever, cuVS vector search | 1 | Live |
| Quote drafting | Writes the quote from calculated prices; the AI writes only the cover note | Kaleido code, Gemini Flash | Open model on NIM | 1 | Live |
| Safety | Flags instructions hidden in uploaded documents, removes invented brands, unsourced links and leaks | Rule-based checks | NeMo Guardrails | 1 | Live |
| Evaluation | Tests every release against a held-back set | Kaleido test set | NeMo Evaluator | 1 | Live |
| Market research | Replaces today’s web research with sourced price bands | — | AI-Q Blueprint | 2 | Planned |
| BOM vision | Finds components from a photo or drawing | — | Multimodal open model on NIM | 2 | Planned |
| HS classification | Assigns HS codes to line items | — | Small open model, NeMo Customizer | 2 | Planned |
| Quote anomaly | Flags overpriced or unusual vendor quotes | — | RAPIDS cuDF, cuML | 3 | Planned |
| Supplier allocation | Splits orders across suppliers at lowest landed cost | — | cuOpt | 3 | Planned |
| Voice intake | Replaces today’s Chrome-and-Edge-only voice input | — | Riva or Nemotron Speech | 3 | Planned |