Live stats
Polled from /api/stats every 5s
Stack
Every entry here is wired into running code on this page, not a bullet list — hover the "how" line for the actual mechanism.
Agentic AI system (LangGraph router → tools → synthesizer)
MCP server patterns (same tools, dual-served)
Ray-parallelized workflows (eval + canary + optimizer fan-out)
Multi-backend runner (Anthropic + Bedrock, swappable)
Mixture-of-experts agent (dispatcher gates 3 scoped experts)
Prompt optimization (propose → score → adopt; 50% → 88% on a real run)
Retrieval-augmented generation (chunk → embed → vector search over real filings)
Adversarial user-simulation testing (LLM personas + critic find real bugs)
Pythonthe entire app — agent, tools, API, training, IaC glue
BashAirflow's BashOperator invokes finagent canaryacross environments- LangGraphthe router/tools/synthesizer agent graph
LangChain@tooldecorator, message types, LangFuse callback integration — shared plumbing across both agents
FastAPI/api/ask,/api/redteam/*,/api/stats, and the static web UI
PyTorchfinagent train— real SFT loop on judge-graded transcripts
sentence-transformerslocal embedding model for filing_search— no API key, runs on CPU
- Rayparallel eval/canary case execution
Kubernetes / EKSJsonnet-rendered manifests, deployed & verified on a local kindcluster
AirflowDAG schedules the nightly canary job
Jsonnettemplates the k8s manifests for local/prod
TerraformEKS/IAM/SageMaker/S3, validate-clean, unapplied (no AWS account)
Dockerapp image, built and run locally- Chromapersistent local vector database backing
filing_search
Git / GitHub ActionsCI: lint + full test suite on every push
AWS (Bedrock, SageMaker)Bedrock model backend + SageMaker training role, real code, no live account
LangFusetraces every LLM call, tagged by environment
Sentryexception capture, wired into CLI + web
Prometheus/metrics— the live stats above come from here; 4 alert rules fire through Alertmanager + Pushgateway
Grafana8-panel dashboard on Prometheus, verified with real data
Weights & Biasesevery eval + optimizer run logged with pass rate + git SHA
Request pipeline
A LangGraph state machine: every question flows through these nodes.
Retrieval pipeline
filing_search, on first use for a given company + form type: fetches the real filing document (not just its metadata), then chunks, embeds, and stores it locally.
Tools
The same functions back the agent and the standalone MCP server.
→ SEC EDGAR (data.sec.gov)
→ Yahoo Finance (yfinance)
→ Yahoo Finance (yfinance)
→ local vector index (Chroma + sentence-transformers) over real SEC filing text
→ Yahoo Finance (yfinance)
→ Yahoo Finance (yfinance)
Observability
How we keep track of what the app is doing, so problems get caught instead of going unnoticed.
Keeps a step-by-step record of how the AI arrived at each answer — like a flight recorder, so anyone can retrace exactly what it did and why.
Keeps score of how the app is performing: how many questions it's answered, how fast, and how often something goes wrong. The live numbers below come from here.
Automatically flags it when something breaks, so issues get noticed and fixed quickly instead of failing silently.
Keeps a report card of how accurate the AI's answers are over time, so we can tell whether a change made it better or worse before anyone else sees it.