LiveFixed income · Markets analytics

CurveIQ

Builder · Full-stack & data · 2026

Descriptive, retrospective yield-curve analytics over US and Indian government rates — term structure, slope/spreads, regimes, real-vs-nominal, PCA and the equity–yield link, each explained in plain language. Strictly descriptive: never advice, never a forecast.

The problem

Yield-curve analysis is powerful but opaque — slope, spreads, real yields and PCA factors are hard to read without a narrator, and most tools blur the line between describing the data and giving advice.

Approach

Two-layer split: L1 is exact, unit-tested, LLM-free compute that produces every number; L2 is a RAG explainer that narrates only the numbers it is given.
A forbidden-language lint blocks any prescriptive or forward-looking output — the system stays strictly descriptive.
Authentic data only (FRED, RBI, OECD, NSE) with a provenance string on every fact row; a daily GitHub Actions refresh runs 40+ validation gates.
Honest US/India asymmetry: India’s missing panels (full curve, breakevens, official recessions) are shown as absent, each annotated with why.

What it does

/us — curve time-scrubber, slope/spreads, real-vs-nominal, PCA factors, equity–yield regime split and crisis overlays (2008/2013/2020).
/in — 10Y, slope, administered repo, Nifty correlation, crisis trajectories, plus annotated ‘missing’ panels.
/calculator — single-bond price↔yield, accrued, duration, convexity, DV01 (US Treasury + India G-Sec presets).
‘Explain’ on every panel: pgvector retrieval over a curated corpus + Claude, grounded in the L1 numbers.

Outcome & results

Bond engine in TypeScript parity-checked against a Python twin via golden vectors.
PCA level factor ≈ 87% of variance; regime-split equity–yield correlation.
Self-updating: a daily refresh keeps data current with hourly ISR — no redeploy needed.

Python

Next.js

TypeScript

Recharts

Claude

OpenAI embeddings

Supabase / pgvector

FRED · RBI · NSE

Live Code Inception

Approach

Two-layer split: L1 is exact, unit-tested, LLM-free compute that produces every number; L2 is a RAG explainer that narrates only the numbers it is given.
A forbidden-language lint blocks any prescriptive or forward-looking output — the system stays strictly descriptive.
Authentic data only (FRED, RBI, OECD, NSE) with a provenance string on every fact row; a daily GitHub Actions refresh runs 40+ validation gates.
Honest US/India asymmetry: India’s missing panels (full curve, breakevens, official recessions) are shown as absent, each annotated with why.

What it does

/us — curve time-scrubber, slope/spreads, real-vs-nominal, PCA factors, equity–yield regime split and crisis overlays (2008/2013/2020).
/in — 10Y, slope, administered repo, Nifty correlation, crisis trajectories, plus annotated ‘missing’ panels.
/calculator — single-bond price↔yield, accrued, duration, convexity, DV01 (US Treasury + India G-Sec presets).
‘Explain’ on every panel: pgvector retrieval over a curated corpus + Claude, grounded in the L1 numbers.