Autonomous AI for Project Intelligence
A serverless system that reads project data, writes standardized AI summaries and status, identifies key participants, and updates the project tracker, with no human in the loop.
Role: Sole technical owner, designed, built, and operated
The problem
Across a large project portfolio, keeping every project's status, summary, and key-participant list current is constant manual work. Updates lag, formats vary by author, and leadership lacks a consistent, at-a-glance view. The information exists, scattered across project records and updates, but nobody has time to distill it uniformly across hundreds of projects.
What I built
I designed, built, and operate a fully autonomous, serverless AI system that does this distillation on its own, on a schedule, with no human intervention. On a daily trigger it reads project data, sends it to a large language model (Amazon Bedrock with Claude) under a structured prompt that returns a consistent format, project overview, latest status, key participants, and writes the results back into the project tracker.
flowchart LR CW(["CloudWatch
daily trigger"]) --> SF["AWS Step Functions
orchestration · parallel · auto-retry"] SF --> R["Lambda
read project data"] R --> AI["Amazon Bedrock + Claude
structured: overview · status · participants"] AI --> W["Lambda
write results to project tracker"]
A scheduled, self-healing serverless pipeline, no human in the loop.
- Orchestration: AWS Step Functions processes many projects in parallel with automatic retry for resilience.
- Structured prompting: the model's output always returns in the same labeled sections, so downstream consumers can rely on it.
- Infrastructure as code: the whole system is defined in AWS CloudFormation, with least-privilege access controls and managed secrets.
It has run continuously in production for approximately a year. As its sole technical owner and on-call engineer, I personally diagnose and resolve every incident, most recently migrating the inference layer to a current Claude model the same day an earlier model it relied on was deprecated.
Why it matters
Twelve months of unbroken autonomous operation, including surviving a model-deprecation cycle, is a concrete demonstration that an AI system can own a recurring knowledge-work task end to end in production, not just in a demo. The pattern (scheduled, structured, self-healing AI summarization writing back into a system of record) transfers to any organization that needs consistent, low-effort status intelligence at scale.