Veera Chunduri

Problem

Insurance and financial advisors piece together a CRM, a handful of calculators, email/SMS tools, and a public-facing site from separate vendors. Nothing serves the whole practice, and nothing is built for how licensing and compliance actually work in that industry.

Decisions

Multi-tenant by tenant_id on every table rather than schema-per-tenant, so any one advisor's data can be extracted or migrated cleanly. Raw parameterized SQL instead of an ORM — at 50 tables, control over the exact query mattered more than the convenience layer. Two fully separate authentication systems: JWT for advisors, TOTP for the internal admin console.

Before treating anything as launch-ready, ran a full self-directed security audit of the codebase — authentication, tenant isolation, input handling, secrets management — and worked the findings into a prioritized fix list.

Outcome

Live in production with a matching staging environment, a written deployment runbook, and a recovery procedure. Six pricing tiers with feature gating enforced at both the API and the UI. 21 CI workflows run tests, security checks, and deploys automatically on every push.

Problem

A working practice needed a public site, online booking, invoicing, document handling, and a blog/content engine that could be run without ever touching code.

Decisions

Split into three independently deployable pieces — marketing site, client portal, API — sharing one database, with billing integration and a full content CMS wired in. When the production host turned out to have no SSH or process-manager access at all, diagnosed that mid-build and rebuilt the deploy path around the host's own Git integration — including catching a GitHub Actions workflow that had been silently failing for a full day while the real deploy path kept working the whole time.

Outcome

Live in production and still actively maintained — the most recently touched project in the portfolio. Recent changes are proactive security fixes (an open redirect closed, dependencies patched) shipped without being asked.

Problem

Most retail strategy backtests are one curve-fit run away from meaningless. A strategy needs to survive out-of-sample data, randomized trade order, and the exact drawdown mechanics of a real funded account before it's worth risking capital on.

Decisions

Built a 7-tier pipeline — in/out-of-sample split, walk-forward optimization, Monte Carlo trade-shuffle, deflated and probabilistic Sharpe ratio, and a full simulator of prop-firm trailing-drawdown rules — with a personal bar above the pipeline's own defaults before anything graduates from backtest to paper trading.

Outcome

One strategy promoted to "validated" after clearing a 0.865 out-of-sample Sharpe across 20 years of data; a competing approach correctly retired after failing out-of-sample despite looking fine in-sample.