Grace AI Platform
AI PLATFORM

Grace AI Platform

Production deployed

The client’s research team was drowning in information. They had data in Oracle databases, research from web sources, and documents from multiple APIs — but no unified way to search, validate, or expand their knowledge base intelligently.

They were spending 6+ hours daily on manual research tasks that should have taken minutes. The core problem was not a lack of data — it was a lack of trust in the data. They needed AI that could tell them not just what it found, but how reliable it was.

We built a trust-aware knowledge expansion engine — a system that does not just retrieve information but scores it for credibility before surfacing it to the user. Every result is tagged with a trust score based on source authority, recency, and cross-reference validation.

The platform integrates Oracle DB (their existing data), Kimi Moonshot API (for long-context AI reasoning), and real-time web search — unified in a single interface where researchers can query in natural language.

  1. 1

    RAG pipeline with pgvector

    Semantic search over Oracle DB content without moving data out of Oracle.

  2. 2

    Kimi Moonshot API

    Chosen for 128K context window needed for long research documents.

  3. 3

    Trust scoring algorithm

    Custom Python module that weights source authority, publication date, and cross-reference count.

  4. 4

    FastAPI streaming

    Responses stream token by token so researchers see results in real-time.

  5. 5

    LangChain orchestration

    Manages the multi-source query routing and result aggregation.

Client

Knowledge management company

Industry

AI / Research

Timeline

10 weeks

Team

Parth + specialist engineers

Status

Production deployed

Tech Stack

LangChainOracle DBRAGFastAPIpgvectorKimi API

Key Results

70% Reduction in research time
3 Data sources unified
128K Context window per query
0%

Reduction in research time

0

Data sources unified

0K

Context window per query

Oracle DB integration was more complex than expected — Oracle’s vector extension was not available on their version, so we built a bridge layer using pgvector as the vector store while keeping Oracle as the source of truth.