How your AI hedge fund actually works.

When you submit a thesis, three fan-out calls run in parallel before the LLM is invoked:

• FRED India macro snapshot — India CPI, RBI policy repo rate, INR/USD, real GDP, 10Y G-Sec yield, global crude (Brent/WTI) — latest observation per series.
• Firecrawl, biased to Indian sources — queries are prefixed with site:nseindia.com OR site:moneycontrol.com OR site:rbi.org.in etc., then top results are scraped and excerpted. Two results, 500-char excerpts — deliberately cheap.
• Strategy Graph similarity — your thesis is embedded with OpenAI text-embedding-3-small and cosine-similarity searched against every past Attevia run. The top-5 nearest are handed to the Researcher as prior art.

Every span is traced in Langfuse, so every memo has a full audit trail of inputs, outputs, tokens and latency.

A Claude Haiku 4.5 call translates the free-form thesis into a falsifiable hypothesis, a named India macro regime, and a universe of 3–8 tickers drawn only from the curated list. The list is India-first: NSE ETFs (NIFTYBEES, JUNIORBEES, BANKBEES, GOLDBEES, LIQUIDBEES, ITBEES, PHARMABEES, CPSEETF, etc.), then G-Secs, then US sector ETFs for cross-border theses.

Optional mode ATTEVIA_AUTORESEARCH_STYLE=1: we borrow the research loop framing from Andrej Karpathy's autoresearch — fixed tool budget, explicit success criterion (evidence-backed hypothesis), keep/discard honesty when data contradicts the thesis, and a simplicity bias. That project runs LLM agents on GPU nanochat training; Attevia maps the same discipline onto markets (no train.py, no val_bpb).

The Quant picks an allocation method — equal_weight, momentum, risk_parity, or regime_blend — a rebalance cadence, and a benchmark (default NIFTYBEES.NS). It never picks weights directly; weights come from the backtest engine.

for each rebalance t:
  momentum_t  ← avg 63-day return across risky assets
  vol_t       ← avg 63-day realised vol across risky assets
  medianVol_t ← median of 22-day rolling vol over past 252 days
  riskOn_t    ← momentum_t > 0 AND vol_t < medianVol_t
  riskyShare  ← 0.80 if riskOn_t else 0.40
  withinRisky ← inverse-volatility weights (risk parity)
  safeSleeve  ← 1 - riskyShare  (gilts / liquid / short bonds)

A Claude Haiku 4.5 call reviews the strategy and lists historical regimes where it would have failed: Mar 2020 COVID, May 2013 taper tantrum, Nov 2016 demonetisation, Feb 2022 energy spike. The Critic's concerns are surfaced on the live trace and in the memo under "Risks".

We fetch up to ~7 years of daily close prices from Yahoo Finance v8 (free, keyless) — whatever liquid overlapping history the universe actually has, with a 60-trading-day minimum. We compute log returns net of a turnover transaction cost, apply the allocation method at each rebalance, and produce: total return, annualised return, annualised volatility, Sharpe (rf = 6.5% for INR, 5% for USD), Sortino, max drawdown, CVaR 95%, win rate. When a universe has no usable price feed, the memo is clearly labelled model-estimated and the chart is an illustrative simulation, not a backtest. The benchmark curve is overlaid for comparison.

Claude Opus 4.8 composes an institutional-style memo: executive summary, key findings, risks, citations. It is forbidden from inventing metrics — every number in the memo must appear on the backtest page, and this is checked against the JSON output. The same system prompt is replayed across runs at 10% of input cost via Anthropic prompt caching, so re-runs are cheap.

The finished memo is serialised into a canonical JSON — run id, createdAt, sha256 of the thesis, strategy spec, sorted weights, and all backtest metrics — and signed with an Ed25519 key that never leaves the server. The signature and public-key fingerprint are attached to every public memo. Anyone can re-fetch the memo, reconstruct the canonical message, and verify the signature with the published public key at /api/pubkey. This is how you know a memo was not altered after publication.