Stack

Multi-agent orchestration with Send() fan-out semantics and custom dict-keyed reducers. Six sector teams + macro economist + CIO + LLM-as-judge run as graph nodes with typed state. Vanilla LangChain doesn't compose for parallel multi-team coordination with state merging.

Haiku for per-ticker quant, qual, and peer-review (~12 calls per Saturday run); Sonnet only for synthesis (macro economist, CIO batch evaluation, nuanced judge cases). Sonnet is ~5× Haiku — reserving it for synthesis keeps per-run cost stable.

Auto-tracing on every production LLM call. The trajectory validator polls LangSmith for graph-correctness invariants (required nodes present, sector_team_node appears exactly 6 times) post-graph.invoke(). Free tier covers a personal-scale workload.

Typed agent outputs (with_structured_output(..., include_raw=True) with strict-mode parse-error contract across every LLM-output site).

Feature store for ~50 features × ~900 tickers × 10y. Migration was driven by a real performance bottleneck: per-feature parquet scans on S3 were slow on the read side and OOMed c5.large training instances. ArcticDB's symbol-keyed range queries pull only the slice each consumer needs.

RAG corpus for SEC filings, 8-Ks, earnings transcripts, and rolling investment theses. HNSW indexing for fast vector search; serverless tier means no idle compute cost when the weekly research run isn't firing.

512-dimensional embeddings tuned for financial-domain text. Cheaper than OpenAI's text-embedding-3-* tier with better fit on the SEC-filing corpus the qual agents retrieve over.

Primary market-data source for adjusted OHLCV and intraday data; yfinance retained as cross-validation only, not as a silent fallback (silent-fallback patterns have been a repeat offender for masking real failures and were retired across the data layer).

Macro indicators and supplemental fundamentals.

Historical portfolio simulation in the backtester.

Primary language across all eight repos. Lambda runs 3.12; local development on 3.13.

Layer-1 gradient-boosted models in the predictor meta-ensemble (momentum + volatility).

Feature engineering, signal scoring, metric computation.

Three pipelines (Saturday weekly research, weekday morning, EOD post-close) orchestrated as state machines. Native to the AWS account that already runs Lambda + EC2 + S3, so no new infrastructure to operate. Execution history is itself the audit trail.

Stateless agent calls, predictor inference, LLM-as-judge, rationale clustering, replay-concordance, replay-counterfactual. Container-image deploys via ECR for the Python 3.12 research Lambda.

t3.small for the executor daemon (stateful, holds the IB Gateway connection), t3.micro for the dashboard, c5.large spot for batch (predictor training + backtester). Right-sized per workload — when predictor_data_prep OOMed on c5.large, the response was a pandas refactor (~1.1 GB → ~91 MB resident), not a bump to xlarge.

Infrastructure-as-Code for Step Functions, Lambda functions, IAM roles, EventBridge rules. Drift detector compares CloudFormation stamps against live AWS state weekly.

Static-site generator for the apex marketing surface — zero JS shipped by default, fast TTFB, build-time SEO. The harness identity benefits from a marketing surface that loads as instantly as the instrument it describes.

Live data console at live.nousergon.ai and the private interview-demo console at console.nousergon.ai. Pragmatic pick for read-only monitoring: a custom React dashboard would be weeks of work for the same outcome.

Paper-account broker connection on the executor host. Hard runtime check refuses to connect to non-paper accounts (account ID must start with "D") — defense against accidental live-account wiring.