💡 Inspiration: The 2:00 AM Problem

Every data engineer knows this nightmare:

  1. At 2:00 AM, an external API silently renames a single JSON key (e.g., author becomes owner_name).
  2. The ingestion pipeline violently crashes. PagerDuty alarms go off.
  3. Downstream dashboards and ML models freeze. An engineer wakes up in the middle of the night just to write a 2-line regex fix.

Why should production pipelines be so fragile?

We built SchemaSentinel-Strands to turn pipelines from brittle glass into self-healing infrastructure: a 24/7 AI Reliability Engineer that intercepts breaking schema drift and repairs data in-flight without dropping a single record.


🛡️ What It Does: 5-Stage Autonomous Healing

When breaking schema drift hits the warehouse, SchemaSentinel heals it in <400ms through 5 visual steps:

  1. 🚨 Intercept: A malformed record fails the strict SQLite schema contract. Instead of crashing, the pipeline catches the error and isolates the payload.
  2. 🧠 Reason: The Strands Agent inspects the failing JSON, compares it with the warehouse schema, and understands column intent using semantic scoring.
  3. ⚡ Synthesize: Powered by AWS Bedrock Mantle (xai.grok-4.6), the agent writes a pure Python transformation patch (transform_record()) in real time.
  4. 🔒 Secure AST Sandbox: The code is compiled in an isolated Python Abstract Syntax Tree (AST) sandbox. Any unsafe call (os, sys, eval) is blocked instantly.
  5. ✨ Conformed Ingress: Numbers, nested arrays, and aliases are normalized to a 0–100 scale and committed to SQLite with a 100% Reliability SLA.

⚙️ How I Built It

  • Autonomous Agent Brain: Strands Agents SDK invoking AWS Bedrock Mantle (xai.grok-4.6) for sub-second code generation.
  • AST Security Barrier: Python ast syntax parser enforcing strict memory and namespace isolation.
  • Chaos Ingress Engine: Real-time multi-partition stream generator simulating real-world upstream API drift.
  • SRE Command Center: A modern Streamlit dashboard featuring live DAG stage nodes, real-time telemetry logs, JSON Before/After diffs, and an interactive Judge Playground.

🧗 Challenges & Breakthroughs

  • No Hardcoded Rules: We eliminated static string checks. The agent uses recursive leaf scoring to differentiate a repo star count (47k) from an actual project rating (4.6/5).
  • Zero-Trust Security: Executing LLM-generated code live in production is dangerous. Our AST sandbox inspects code structure before compilation, with an instant deterministic fallback if compilation fails.
  • Sub-Second Speed: Optimized prompts and Bedrock Mantle inference achieve full drift-detection to warehouse-commit in ~380ms.

🧠 Key Takeaway

Agents shouldn't just be conversational chatbots. In SchemaSentinel, the agent acts as active, self-healing system middleware—writing and executing micro-patches at runtime to eliminate infrastructure downtime.


🚀 What's Next

  • Native connectors for Snowflake, BigQuery, and Databricks.
  • Automated GitHub Pull Request generation to update upstream dbt models permanently.
  • Production Kafka / Apache Flink streaming consumer integration.

Built With

  • ast
  • autonomous-agents
  • aws-bedrock
  • aws-bedrock-mantle
  • data-engineering
  • etl
  • grok-4.6
  • pandas
  • python
  • react
  • rest-api
  • sqlite
  • strands-agents-sdk
  • streamlit
  • tailwind-css
  • tavily-api
  • typescript
Share this project:

Updates

Submission history