Inspiration
Modern IoT frameworks generate massive streams of real-time telemetry data. However, edge sensors are notoriously prone to hardware degradation, communication drops, and environmental interference, leading to corrupted records, missing vectors, and extreme mathematical outliers. Traditional Rule-Based Software Reliability Engineering (SRE) systems are rigid, failing to catch multi-variable anomalies that fall outside hardcoded thresholds.
We were inspired to build TelemetryGuard Agent to bridge the gap between deep analytical reasoning and automated data operations. By utilizing Agentic AI, we wanted to move past passive alert systems and build a self-healing database pipeline that autonomously reads, reasons about, and corrects data stream anomalies under structured human oversight.
What it does
TelemetryGuard Agent is an autonomous SRE data auditor that continuously monitors MongoDB IoT database collections. Instead of just flagging errors textually, the agent evaluates incoming document feeds, runs multi-step statistical sanity checks, and takes real-world corrective actions:
Scans & Filters: Pulls unverified records from raw telemetry data collections.
Evaluates Integrity: Uses advanced reasoning to detect missing key-value pairs, structural schema drift, and physical sensor outliers.
Executes Remediation: Autonomously modifies the state of anomalous records to isolate them from downstream production analytics and writes structured hardware maintenance tickets into a separate system database.
How we built it
The core intelligence engine is powered by Gemini, orchestrated seamlessly through Google Cloud Vertex AI Agent Builder.
Data Layer: Built using a distributed MongoDB Atlas cluster storing mock high-frequency environmental sensor payloads.
Tool Orchestration: Integrated the official MongoDB Model Context Protocol (MCP) server, exposing core database operations (find, updateOne, insertOne) to the agent's runtime environment.
Network Gateway: Securely mapped the local MCP infrastructure to the Google Cloud environment using a tunneling gateway, allowing the LLM to dynamically generate structural database queries based on real-time observations.
Challenges we ran into
Building an autonomous system within a tight hackathon timeline introduced distinct engineering bottlenecks:
Tool-Calling Latency: Setting up a low-latency execution loop between the cloud agent builder and localized database drivers required precise performance tuning of JSON-RPC payloads over the Model Context Protocol.
Context Grounding: Forcing the LLM to strictly adhere to deterministic mathematical boundaries rather than hallucinating data corrections required intensive system instruction design and explicit few-shot engineering.
Asynchronous State Alignment: Ensuring that the agent correctly tracked modified document states without infinite looping through the same data pool required strict status flag updates in the MongoDB document schemas.
Accomplishments that we're proud of
Successful deployment of a functional, end-to-end autonomous agent loop using Google Cloud Agent Builder and MongoDB MCP within a 26-hour sprint.
Constructing a rigid reasoning protocol where the agent safely translates open-ended observational logic into highly deterministic database operations without corrupting existing records.
What we learned
This project provided immense clarity on the architectural differences between traditional deep learning pipelines and Agentic workflows. While deep learning excels at feature extraction and pattern prediction, Agentic AI introduces a dynamic execution loop:$$\text{Reasoning} \rightarrow \text{Planning} \rightarrow \text{Tool Execution} \rightarrow \text{Observation}$$We gained hands-on experience in implementing the Model Context Protocol (MCP), understanding its power as an open standard for connecting intelligence models to isolated enterprise data layers.
What's next for TelemetryGuard Agent
Deep Statistical Modeling: Integrating dedicated statistical outlier detection algorithms, evaluating if an item falls outside a dynamic threshold defined by z-scores:$$Z = \frac{x - \mu}{\sigma}$$Where $x$ is the data point, $\mu$ is the rolling mean, and $\sigma$ is the standard deviation, using the agent to orchestrate the mathematical computation before modifying records.Production Native Deployment: Migrating from a tunneled local gateway to a fully containerized, serverless architecture on Google Cloud Run to ensure secure, auto-scaling enterprise deployment.
Log in or sign up for Devpost to join the conversation.