# I built an agent that reads SEC filings so I don't have to > Source: > Published: 2026-08-28 16:57:15+00:00 Every public company's story is hiding in plain sight — in 10-Ks and 10-Qs that almost nobody reads end-to-end. The good stuff is specific: a gross margin inflecting, risk-factor language that wasn't there last quarter, a going-concern sentence buried on page 60. I wanted that surfaced to me every morning without me doing the reading. So for the **All Things Agentic Hackathon** I built **EDGAR Sentinel**: an autonomous agent on Google Cloud that wakes up at 6:30 every morning, scans SEC EDGAR for new filings across a 30-company watchlist, reads them with a two-model pipeline, remembers every prior filing, and emails me what changed — with a public dashboard for everything it knows. Cloud Scheduler → Cloud Run Job → an **ADK orchestrator agent** (Gemini 3.5) whose tools are the pipeline stages → SEC EDGAR (politely: declared User-Agent, throttled) → raw filings archived to Cloud Storage → a section parser → **Gemma** (on its own Cloud Run service, via Ollama) writes triage notes → **Gemini 3.5 on Vertex AI** scores the filing against a five-pillar "Filing Health Score" with schema-enforced JSON → Firestore stores it → a **delta engine** compares against the company's prior filing and fires deterministic One design rule shaped everything: **agentic control flow, deterministic execution**. The LLM decides Gemini 3.5 Flash does the deep reading: five pillar scores with cited rationale, extracted metrics, three decision-relevant highlights. Temperature zero, pydantic schema enforced, composite recomputed in code so config — not the model's arithmetic — is authoritative. Gemma's job is deliberately smaller: read the risk-factors section and produce a dozen terse triage bullets — red flags, notable changes, tone — that ride along to Gemini as a second opinion. It runs scale-to-zero on CPU. My first design had Gemma *rewriting* filing text; that was wrong in an instructive way (below). `done_reason: length` — the model spent its whole output budget on hidden reasoning and never wrote the answer. One `think: false` later, 33-second useful triage notes.`RIS K FACTORS` , and repeats "Item 1A" as a page header through the whole section — my "take the last heading match" heuristic found nothing. Fix: match headings with optional intra-word whitespace and take the match with the `make_client().models.generate_content(...)` ) let the client get garbage-collected mid-request; its finalizer closed the HTTP pool: `Cannot send a request, as the client has been closed.` Cached singleton.`constraints/vertexai.allowedModels` ) and strips default service-account grants — both showed up as cryptic 400s/403s. Both fixed with scoped, least-privilege IAM rather than hammer-sized grants.Every one of these would have detonated during a live demo. Finding them on day one and day four instead is most of what "production-minded" means. The delta engine is the feature I'd defend in a knife fight. Because every analysis persists in Firestore, each new filing is compared with the company's prior one — pillar by pillar — and a deterministic rule (≥10-point move, band change, or risk-pillar collapse) decides whether to alert. On the full backfill it flagged, among others: Plug Power sliding **Caution → Distress** (cash down to $161.9M), Salesforce and Meta dropping out of Strong, and Coinbase and AMC genuinely recovering. It also caught Apple's management going cautious on component costs a quarter before it showed up anywhere else in the filing — a 10-point management-signal drop while the composite barely moved. 30 companies · 58 filings analyzed · 8 live alerts · running unattended every morning since August 14 · ~1 minute per filing · 20 unit tests · roughly a dollar a day in cloud costs while idle-scaling to zero. One last production note: even the demo video is Google AI — narration by Cloud Text-to-Speech (Chirp3-HD), soundtrack generated with Lyria 2 on Vertex AI, and the screen captured while the real daily job ran live on Cloud Run. *I created this piece of content for the purposes of entering the All Things Agentic Hackathon.* EDGAR Sentinel produces automated research summaries