An agent that reads SEC filings so nobody has to.
Every morning EDGAR Sentinel found new 10-K and 10-Q filings, read them with a two-model pipeline, compared each one with the company's previous filing, and emailed what changed.
We built it in three weeks for Google Cloud's All Things Agentic Hackathon and left it running on the same stack we build for clients. It ran every morning from August 14 to October 28, 2026, and we retired it once judging closed. The code, the demo, and the write-up stay public.
Every quarter, public companies file 10-Qs and 10-Ks, dense documents almost nobody reads end to end. The signal is specific and buried: a margin turning, new risk-factor language, a going-concern sentence on page sixty. Most businesses have a version of this pile, and the same pattern applies to contracts, inspection reports, and regulatory bulletins.
What it did at 6:30 Pacific, with nobody watching.
Scan
Cloud Scheduler started a Cloud Run job. An orchestrator agent built on Google's Agent Development Kit checked SEC EDGAR for new filings from a thirty-company watchlist, identifying itself and respecting the rate limit, then archived the raw documents to Cloud Storage.
Read
A parser sliced each filing into its sections, surviving the quirks of inline XBRL. Gemma, a small open model on a private Cloud Run service, wrote terse triage notes on the risk factors. Gemini 3.5 on Vertex AI then scored the filing on five pillars with cited reasons, as schema-enforced JSON at temperature zero.
Remember and compare
Every analysis persisted in Firestore. Each new filing was compared with the company's previous one, pillar by pillar. A threshold rule you can read decided whether that was an alert, and Gemini wrote the short “what changed” narrative.
Deliver
An email digest, alerts first. A public read-only dashboard with health bands, trend lines, and a drill-down that showed every pillar's reasoning, the quarter-over-quarter change, and Gemma's notes.
The same habits we bring to client work.
It held up outside the demo
The schedule fired every morning from August 14 to October 28 with nobody watching: 93 runs, 92 of them clean. Each filing was processed in isolation, so one bad document was logged and skipped instead of stopping the run.
Every answer was auditable
Scores were five pillars with cited reasons, recomputed in code. Alerts came from a threshold you can read, not from how the model felt that day.
Least privilege, by default
Three service identities, each with only what it needed. The one secret lived in Secret Manager. Nothing sensitive in the code, the image, or the repository.
Under four minutes, live on Google Cloud.
The demo starts the job on camera and comes back for it finishing: the Cloud Run execution, the console, the digest arriving, the dashboard updating.
EDGAR Sentinel produced automated research summaries from public SEC filings. It was not investment advice. Built for Google Cloud's All Things Agentic Hackathon and submitted on behalf of ASI. Judging closed in October 2026, and the system was retired on October 28.
The same pattern, on your documents.
Contracts, inspection reports, supplier bulletins, regulatory notices: anything that changes quietly and matters when it does. Bring the pile and we will tell you what it would take.
Prefer email? contact@asisystems.io