FIG. B · Door B · Credit risk, market analytics, lender research
Markets & Risk
This practice is held to a stricter rule than most portfolios allow: every number must be reproducible from committed artifacts. The scorecard below runs the repository’s own lookup table and policy YAML; the backtest charts were produced by running the repository’s own engine over its own data lake; the lender leaderboard is extracted data, not a retyped claim. If a number can’t be traced, it isn’t here.
FIG. B.1
Lending Lifecycle Platform
300k+ Home Credit applications into a WoE scorecard, a YAML policy engine with adverse-action codes, an auditable decision API, and a simulated 500k-loan book.
KS 0.297 · GINI 0.404 · PSI MONITORED
FIG. B.2
Market-Data Ops & Portfolio Analytics
A cron-fed Parquet lake of NSE bhavcopies, VIX, NAVs and FX — reconciliation-grade — with VaR, drawdowns and a bootstrap-tested momentum backtest.
SHARPE 0.93 (RUN) · 2018–25
FIG. B.3
Listed-Lender Equity Research Workbench
Twelve listed lenders' disclosures extracted into source-cited ratio families, driver-based models, and timestamped initiation notes tracked against Nifty Bank.
12 LENDERS · COMPOSITE QUALITY RANKING
FIG. B.4
Banking-Documents GenAI Workbench
Bank-statement parsing into FOIR-grade features, regex-plus-Presidio PII masking you can test live here, and cited RAG over RBI Master Directions.
93% FIELD ACCURACY · 94% CITATION FIDELITY
FIG. B·W · The desk months — analytics lens
Work, read as facets
Mar – Apr 2026 · Remote
Financial Analyst Intern — AQUA Pvt. Ltd.
Two months on a trading desk, doing the analytics the desk actually runs on. Intraday OHLCV across 40 NSE scrips (90 days, Python and SQL) flagged trend breaks and volatility regimes; nine mean-reversion and momentum rules were backtested over two years of OHLCV, the best at post-cost Sharpe 1.6 / 8% maximum drawdown. Daily derivatives screens over 180 NSE symbols fed 8–10 setups a day to the intraday desk.
Risk sat beside return: a Python risk sheet computed daily VaR, beta and stop-loss levels for a 20-name book and escalated 3 limit breaches. Out-of-sample threshold experiments lifted the strategy hit rate from 52% to 57% pre-rollout, weekly P&L attribution dropped from four hours to 20 minutes across three books, and the 15-KPI desk MIS became automated EOD dashboards. The fundamental side ran in parallel: 3-statement and DCF-comps models for 14 NSE-listed names, six investment notes, and a 5-stock book shaped by them.
Jun – Jul 2026 · New Delhi
Analytics facet — Founder's Office (Software & Automation) Intern — The Rocket Media
The same analytics habit, pointed at content. A Python ETL unified Substack, three YouTube channels and podcast stats into one reporting layer feeding a weekly Looker Studio dashboard (views, watch-time, retention, CTR). The ML reel-selection funnel was audited in SQL: 15% conversion (11 of 75 candidates), and a failure-mode Pareto retuned the prompt thresholds.
→ the same months through the engineering lens: Door A · Work, read as facets
The first practice lives behind Door A — Engineering & AI. The person behind both is in About.