Finance lens

Axiom shipped first as a typical retail valuation tool: pull a company's financials, run a DCF, blend it with EV/EBITDA and P/E multiples, layer an ML "correction multiplier" on top, and call the output "DCF Fair Value." It wasn't. The displayed fair value could read $531 while the underlying DCF was $380, the gap was ML fudging and silent multiple expansion, and the sensitivity table inside the same app contradicted the headline. So I tore it down and rebuilt around one rule: the number has to mean what it says.

Engineering lens

The blending pipeline, the ML correction multiplier applied to fair value, and the headline-vs-internals inconsistency were all removed in one pass. The DCF became a single deterministic Python function with an audited input contract; comparables were pulled out of the headline number entirely and re-rendered as labelled side context. A HARDCODED_VALUES.md file was added inventorying every remaining hand-set constant in the codebase, with a phased plan to migrate them to live peer-comp estimates, so what's assumed is auditable, and what's computed is real.

Finance lens

The core is a 10-year Discounted Cash Flow, the same instrument a banker builds in Excel for an IPO pitch, but coded as a deterministic pipeline. Free cash flow is projected with a three-stage growth fade (current → industry → terminal), discounted at a WACC built from CAPM equity costs and a synthetic credit-rating spread (Damodaran-style interest-coverage-to-rating mapping, so highly-covered firms automatically get cheaper debt, no hand-typed numbers in the spread table). Net debt comes off enterprise value, divide by shares, done.

Engineering lens

Terminal growth is capped strictly below WACC so Gordon Growth can't blow up. The same DCF math powers both the headline number and the sensitivity table, every figure rendered in the UI is internally consistent. One formula, no blending, no ML on the output.

Finance lens

A vanilla DCF works for Microsoft and dies on Tesla, banks, REITs, distressed names, hyper-growth, and "story stocks" don't share a single valuation language. Axiom routes every ticker through one of 9 archetype frameworks (FINANCIAL, GROWTH, MATURE, CYCLICAL, HIGH_CAPEX, HYPER_GROWTH, TURNAROUND, DISTRESSED, STABLE_GROWTH) which decides the formula: banks get P/B, REITs get P/FFO, mature operators get a clean DCF. On top of that, a 54-tag sub-sector table calibrates the multiples shown alongside.

Engineering lens

Archetype routing maps revenue profile, margin shape, leverage, and sector tags at request time. Each archetype has its own DCF inputs (terminal growth, fade schedule, reinvestment assumptions); financials and REITs route to entirely separate model functions. The 54-tag multiplier table lives in subsector_multiples.json and acts as a Bayesian prior on the comp multiples, not the answer, the starting point.

Finance lens

The original ML layer learned to patch its own past errors, a circular feedback loop that doesn't actually predict the cross-section of returns. The rebuilt version works the way a real quant shop would: it scores every company on three independent factors, Value (DCF upside vs. market), Momentum (Jegadeesh-Titman 12-1 month sector-relative return), and Quality (FCF yield), Z-scored within sub-sector so Google is compared to internet peers, not oil majors. Crucially: this signal is not yet shown to users. It's being measured first.

Engineering lens

Information Coefficient + quintile hit rates are computed walk-forward in ml/monitor.py, the metrics quant funds actually use to evaluate predictive signals, not MAE. A "✓ SKILL SIGNAL" annotation appears in the monitoring output once t-stat ≥ 2 holds across enough OOS periods. The factor model exists in the codebase but is intentionally absent from the user-facing UI until it earns its place there. The model is allowed to say "I don't have a signal yet."

Finance lens

Layperson reading: the app gives "is this stock cheap or expensive" with a number, a colour, and a comparison to Wall Street. Finance reading: a fully-instrumented DCF + multi-factor research pipeline. Same product, two reading levels. The demo site has a dedicated "What this isn't (yet)" section that names the limitations: yfinance can be stale, several constants are still hand-set, the factor model isn't live, US-only coverage. Owning the gaps is part of the product.

Engineering lens

Architecturally a Flask + Postgres app with a vanilla-JS frontend, yfinance as the data spine, and an XGBoost-style walk-forward evaluation harness behind the scenes. Deployable on a single container. Every hand-set constant is named in HARDCODED_VALUES.md. Yahoo balance-sheet ratios known to be unreliable (ROE, ROIC, D/E, Altman Z) are currently hidden from the UI rather than shown wrong, visible-honest beats invisible-broken.