Quality assurance · Production

Zectonal builds a Rust-based AI agentic framework for auditable multimodal data quality monitoring

The problem

Zectonal's rules-based approach to data quality monitoring could not scale as the number of supported data sources and file codecs grew, and Python-based AI agent frameworks proved too complex and unreliable for validating which function tools an LLM actually called.

First attempt

Python-based AI frameworks (LangChain and LlamaIndex) were evaluated but found overly complex, with no reliable mechanism to detect when an LLM called the wrong function tool, leaving human intuition as the only check.

Workflow diagram · grounded in source
1
Continuous data pipeline ingestion
trigger
“analyzing similar files flowing through a data pipeline every few seconds on a 24/7 basis”
2
LLM selects function tools
ai_action
“having LLM's decide what function tools to call instead of our pre-defined rules-based approach”
3
Agents run diagnostic algorithms
ai_action
“specialized AI agents to call specific diagnostic algorithms, or function tools, that can characterize large data sets in order to find and diagnose quality defects or malicious content”
4
Agent Provenance tracking
output
“We call this ability to track agent communications from initial tasking to completion, and everything in between, Agent Provenance, named after the concept of Data Provenance”
5
Real-time UI visibility
output
“we can now provide real-time feedback into our UI so users can see what agents and function tools are called, and track their progress responding to a task or question”
6
Prompt and model optimization loop
feedback_loop
“allowed us to develop better agent prompts, better function tool descriptions, and allowed us to fine tune models more efficiently where needed”
Reported outcome

Zectonal built a Rust-based AI agentic framework that scaled to dozens of specialized agents, introduced Agent Provenance for auditable tracking of all agent communications, and added real-time UI visibility into agent and tool activity.

Reported metrics
Data analysis speedsub-second speed
Specialized agents deployeddozens of specialized agents
Reported stack
RustOpenAIAnthropicOllamaClaude SonnetGpt-4oGpt-4o-minillama.cpp
Source
https://zectonal.medium.com/why-we-built-our-ai-agentic-framework-in-rust-from-the-ground-up-9a3076af8278
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Zectonal built a Rust-based AI agentic framework that scaled to dozens of specialized agents, introduced Agent Provenance for auditable tracking of all agent communications, and added real-time UI visibility into agen…

What tools did this team use?

Rust, OpenAI, Anthropic, Ollama, Claude Sonnet, Gpt-4o, Gpt-4o-mini, llama.cpp.

What results were reported?

Data analysis speed: sub-second speed; Specialized agents deployed: dozens of specialized agents (source-reported, not independently verified).

What failed first in this deployment?

Python-based AI frameworks (LangChain and LlamaIndex) were evaluated but found overly complex, with no reliable mechanism to detect when an LLM called the wrong function tool, leaving human intuition as the only check.

How is this quality assurance AI workflow structured?

Continuous data pipeline ingestion → LLM selects function tools → Agents run diagnostic algorithms → Agent Provenance tracking → Real-time UI visibility → Prompt and model optimization loop.

Related quality assurance cases