Skip to content

Technology

Engineering for honest answers.

Python services (FastAPI, SQLAlchemy, LangGraph) behind a Next.js terminal. Everything below links to code in the repository; nothing is described that does not exist.

Point-in-time data pipeline

Every record is stored with the moment it became public, and every read asks “what was known at time t?”.

  • SEC facts usable from filing date + 1 day; 8-K events from SEC acceptance time
  • FINRA short interest from settlement + 10 days; Treasury from date + 1 day
  • Daily bars never include a still-forming session
  • OpenBB (Apache-2.0) runs in an isolated venv behind OPENBB_DATA_ENABLED

services/market-data

AI research council

A LangGraph graph of fundamental, technical, quantitative, macro and sentiment analysts feeds a CIO synthesis and a portfolio constructor.

  • Default model runs locally (LM Studio, Qwen) — no data leaves the machine
  • Optional Claude backend behind daily/monthly budgets and per-minute rate limits
  • 60-second per-agent budget; untrusted text passes an injection filter
  • Outputs are theses with evidence, never orders

services/intelligence

Research engine

A separate package for asking whether anything works, built to report failure as readily as success.

  • Purged, embargoed walk-forward evaluation; final holdout locked and usable once
  • Stationary-bootstrap confidence intervals, probabilistic & deflated Sharpe, Holm correction
  • Experiment registry that records every trial, including failures
  • Risk-overlay and pod-allocation studies with incremental contribution

services/research

Validation & promotion

Strategies move EXPERIMENTAL → BACKTEST_VALIDATED → PAPER_ELIGIBLE → PAPER_ACTIVE only by a named human.

  • Reproducible result hash (validation runs twice)
  • Costs included and re-run at 2×
  • Benchmarked against an equal-weight universe and SPY
  • Survivorship-bias flag on every result

services/fund/src/hf_fund/promotion.py

Order management (paper)

One OMS, one state machine, one broker path — Alpaca's paper endpoint.

  • Approvals bound to a hash of the exact order payload; 30-minute expiry
  • Risk re-checked at the live price immediately before submission
  • Exact paper-host allowlist checked before every submit
  • SUBMIT_UNKNOWN reconciled by client order ID — never blindly retried

services/execution

Accounting

An append-only ledger (enforced by database triggers) is the financial source of truth.

  • Broker activities (dividends, fees, splits) ingested idempotently
  • Ledger cash reconciled with broker cash; drift halts trading
  • Partial fills applied as deltas from broker truth

packages/py-core/src/hf_core/ledger.py

Observability & audit

A hash-chained audit log records every agent run, proposal, risk decision, approval, submission and fill under a correlation ID.

  • Structured JSON logs; secrets never logged
  • Execution latency telemetry from the audit log
  • Kill switch state in Redis with a durable database record

packages/py-core/src/hf_core/audit.py

Technology · Altunyan AI Hedge Fund