Agentic Data Structuring & Query Platform
An orchestrator/sub-agent pipeline that turns raw multi-source data into tagged, queryable records — and answers questions about it in plain English.
- Context
- Avira Digital Technologies — BMS client
- Period
- 2026 — Present
- My role
- ML Associate — developed the LangGraph sub-agents, the MCP schema/query tool servers, and the governance context integration.
The problem
Raw data arrived from many sources with no consistent structure, which made it unusable for analysis — and even once structured, answering questions about it required someone who could write SQL.
Why it needed solving
Manual tagging and categorization does not scale with ingestion volume, and a structured database that only engineers can query leaves most of its value unrealized. Both halves of the problem had to be solved for the data to be useful to the business.
The approach
A multi-stage agentic pipeline: specialized sub-agents structure and enrich records on the way in, and MCP-exposed schema-introspection and query tools let downstream agents answer natural-language questions by generating SQL against the live schema.
Architecture
- Ingestion stage — raw, multi-source data collected into a common staging shape.
- Structuring stage — topic modeling groups related content before records are categorized.
- Orchestrator/sub-agent layer — 10+ deep agents in LangGraph, each specialized: sentiment analysis, thematic clustering, entity extraction.
- Persistence — enriched, tagged records written to the database.
- MCP query layer — servers exposing schema-introspection and query-execution tools.
- Query stage — downstream agents translate natural-language questions into dynamic SQL against the introspected schema.
- Governance layer — business-logic and governance rules held as structured context files, constraining what agents may answer.
Implementation
- Developed 10+ specialized LangGraph agents under an orchestrator/sub-agent architecture, so each enrichment concern stayed independently testable.
- Split enrichment across sentiment analysis, thematic clustering and entity extraction sub-agents rather than asking one agent to do all three.
- Applied topic modeling to group related content ahead of categorization.
- Built MCP servers exposing schema-introspection and query-execution tools so SQL generation targets the real, current schema.
- Encoded business-logic and governance rules as structured context files so agent answers stay aligned with client data-governance policy.
AI/ML components
Orchestrator/sub-agent architecture in LangGraph
Sentiment analysis sub-agents
Thematic clustering sub-agents
Entity extraction sub-agents
Topic modeling for content grouping
Natural-language-to-SQL generation over introspected schemas
Governance rules as structured agent context
Challenges & how they were handled
A single agent asked to structure, classify and extract at once is hard to debug and harder to improve.
An orchestrator delegating to 10+ narrow sub-agents, each with one responsibility and its own failure mode.
Generated SQL breaks the moment it assumes a schema that has drifted.
MCP schema-introspection tools, so agents read the live schema before generating a query rather than relying on a snapshot.
An agent that can query anything will eventually answer something it should not.
Governance and business-logic rules supplied as structured context, keeping answers inside client policy.
Results
10+
Specialized agents built
What comes next
- Widen sub-agent coverage to additional enrichment dimensions.
- Tighten query-generation evaluation against a labeled question/SQL set.
Related work
Athena — Enterprise Agentic AI Platform
Avira Digital Technologies — internal product
Provider-agnostic LLM gateway, MCP tool servers, layered agent memory, guardrails and an LLM-as-judge evaluation harness.
Clinical Knowledge Graph RAG
Avira Digital Technologies — BMS client
A clinical Q&A retrieval system that moved from hybrid-search RAG to Graph RAG once relationship modeling became the real bottleneck.