All workflows
Research

Portfolio monitoring and signal generation

Recurring agentic run across multiple tickers - news, filings, macro data - producing structured trade signals with compounding history.

Portfolio signal generation on a Salesforce + Domo + Zapier stack: 64% invisible in prior-signal history and connector fan-out. Firedog tracks accumulation drift per ticker across every intraday run.

Model 17%Prompt 19%Orchestration 4%Retrieval 10%History 41%Connectors 8%
ComponentWhy it is unattributedcalls$ / run
Model · 17%Signal inference (GPT-4.1)For each ticker batch, the model receives current news context plus the signal schema and produces a structured JSON signal object.tokens in: 38,000 · tokens out: 4,200Visible on the invoice.15$0.11
Prompt · 19%Signal schema, ticker metadata, and examplesA 4,200-token prompt encoding signal categories, confidence thresholds, and sector-specific few-shot examples is sent on every ticker-batch call.tokens in: 63,000 · tokens out: 0Visible on the invoice.15$0.13
Orchestration · 4%Ticker batching, routing, and aggregation callsA router agent groups tickers into batches, dispatches to signal agents, and an aggregator merges outputs. Each produces a separate billed call.tokens in: 9,500 · tokens out: 1,100Batching and aggregation calls are attributed to 'orchestration infrastructure' in the platform, not to this workflow's cost center.18$0.028
Retrieval · 10%News and filing vector lookups per tickerEach ticker triggers an independent vector query for the last 24 hours of news and any recent filings. Queries are not batched; 150 tickers produce 150 round trips.tokens in: 28,000 · tokens out: 0Vector DB query costs are on the Pinecone or Weaviate invoice; no join to the LLM workflow P&L has been built.150$0.067
History · 41%Prior signal state re-injectionThe last 3 signal outputs per ticker are prepended on every run so the model can detect trend changes. Across 150 tickers this accumulates 90k+ tokens per run.tokens in: 112,000 · tokens out: 0Signal state history is maintained by the application layer as a 'context window' feature; none of the LLM vendor dashboards show it as a distinct cost component.15$0.27
Connectors · 8%Salesforce, Domo, Slack, and Zapier connectorsEvery run writes signal objects to Salesforce, refreshes the Domo dashboard, posts to Slack, and triggers downstream Zaps. Because signals are produced per ticker batch, the write-back and normalization calls fan out across the 150-name universe on each run.tokens in: 8,000 · tokens out: 900Salesforce and Domo writes bill to their platform contracts and each Zapier task to the Zapier plan; the connector volume scales with ticker count but is never charged to the signal workflow.14$0.050
Portfolio signal generation on 150 tickers costs $0.65/run; run it 12x/day for a year and 64% of the $15.6k annual bill is invisible history re-injection and Salesforce/Domo/Zapier connector fan-out across ticker batches.
$0.65
cost per run
$16k
per year
64%
unattributed today
$7.80
per user, per day
Scenario: Mid-size quant fund, 150-ticker universe. 8 active users, 12 runs/day, 250 working days. 45 documents/run. Avg doc: 3,000 tokens.

Default model: gpt-4-1

Trigger: Scheduled: intraday every 30-60 minutes during market hours, plus event-driven on news alerts and filing filings. History compounds across every run.

Who runs it: Systematic and quant companies, multi-team pods, and prop trading desks running intraday or end-of-day signal pipelines across 50-500 names. Signals write back to Salesforce, feed Domo dashboards, route through Slack, and trigger existing Zapier flows, so every intraday run fans out across that connected stack.

What Firedog shows

  • Per-run cost with history accumulation graphed over the trading day - showing cost-per-run drift as history grows.
  • Which tickers are generating the most history volume and driving disproportionate per-batch cost.
  • Projected annual saving from rolling a 1-signal rather than 3-signal lookback per ticker.
  • How connector write-back to Salesforce, Domo, and Zapier scales with the ticker universe and inflates per-run cost invisibly.

Decisions this informs

  • Reduce the signal history window from 3 prior runs to 1 per ticker; test accuracy delta before committing.
  • Batch ticker vector queries into grouped calls; 150 individual round trips is the single largest latency and cost driver.
  • Route ticker batches with no recent news activity to GPT-5 Mini or Claude Haiku 4 to cut per-run model cost by 75%.
  • Write back to Salesforce and Domo once per run as a batched upsert rather than per-ticker, and consolidate the downstream Zaps that duplicate routing.

Connected systems

  • Salesforce · CRM
    Generated signals update opportunity and contract-value fields on covered-name records in Salesforce.
  • Domo · BI / dashboards
    Structured signal objects are pushed to the fund's Domo signal dashboards for PMs.
  • Slack · Collaboration · Salesforce (Slack)
    High-confidence signals are routed to desk Slack channels in real time.
  • Zapier · Automation
    Existing Zaps route signal outputs to downstream trackers; the AI run must cohabit with and extend those flows.

Understand every AI dollar.

Book a demo to see how Firedog attributes every AI request across your organization, and what your leadership team can do with the data.