Associate Director, S&P Global
Fifteen years deep in how software actually works. That's exactly why AI builds fast in my hands.
Named inventor on a published machine-learning patent (S&P Global, 2025) for predicting emerging industries from data.
The timeline runs newest first. Four threads run the whole length of it; tap one to light up its builds.
Deep, opinionated documentation on AI coding tools and agentic engineering, written from daily practice.
Learned: Context engineering: the right context is what closes the gap between an agent's capability and expert-level output.
Building the engine that turns options math into an index: pricing logic, validation against years of historical data, and the correctness checks an index has to survive before anyone trusts it.
Learned: Quant index engineering under correctness and latency constraints.
Prototyped production-grade LLM applications across OpenAI, Gemini, Bedrock and Llama via LlamaIndex and LangChain:
Learned: Fluency across every major model and framework, plus architecting around a model's weakness instead of waiting it out.
The LLM work I run on my own products, started the year ChatGPT changed the math:
Learned: Personal products are the honest benchmark for AI work: nobody sees the demo, only whether the pipeline holds at millions of items.
Built deep-learning and SetFit few-shot classifiers that assign companies to thematic indices from public S-1 filings, beating the classical TF-IDF / logistic-regression baseline on false positives and gaming-resistance. The work behind patent US20250094860A1.
Learned: Few-shot deep learning can beat a hand-tuned classical baseline when data is scarce, and is harder to game.
Migrated 50+ Excel-based index calculations into a production Java/Jython engine. Then built a small purpose-built language (a domain-specific language, with ANTLR) and its UI, so index managers, not engineers, could write calculation logic themselves. Recognized with a promotion; won a company-wide hackathon along the way.
Learned: Compiler tooling from a decade earlier, aimed at a new problem: let domain experts write logic, not tickets.
Started from a decision: I wasn't going to trade manually. Discovered Zerodha exposed an API, opened an account and built against it. When the official API's pricing didn't fit a hobby project, I built an unofficial client on the same endpoints their own platform uses. The bot pulled live quotes and executed configurable strategies.
Learned: How the financial-technology journey began: market mechanics, learned by building against them.
A learning-and-quiz platform I've shipped for a decade. Started on PHP/CodeIgniter, rewritten onto Ruby on Rails + PostgreSQL where the core still runs; the quiz player app has since moved to Elysia + Astro. A live multiplayer quiz product is prototyped alongside. Mine end to end: code, infra, content.
Learned: Carrying one product across full rewrites and ten years without letting it die.
An exam-prep platform: solved question banks organized by branch and semester. Contributed and built the app along with its founder, working across the whole stack.
Learned: Taking a product to real users as a team of two.
Built and led the UI team as a player-coach (team lead and lead developer in one) for the operations portal of Reliance Jio Money (now Jio Financial), and later the ops portal for Jio Payments Bank. Every operator-facing UI workflow, including the maker-checker flow, was built by our team. Ran in production for roughly ten years.
Learned: The first leadership beat: leading the team while staying in the code.
On GE Capital's loan system, streamed millions of records to the browser using Server-Sent Events (non-obvious at the time), and that initiative won the UI scope for the team.
Learned: Initiative past the assigned boundary, and streaming at scale before it was common.
A parser and source-to-source transcompiler that migrated a legacy mainframe system from COBOL to Java. Grammar-driven translation most colleagues didn't think was possible.
Learned: Compilers and language tooling: the start of a code-transformation thread that runs straight to the AI work today.
A habit that started in TCS training: the intranet blocked the game sites, so I built a live multiplayer bingo the whole batch played over the LAN. It spread to other batches, and the program head asked me to stay on and build learning games for new recruits. The habit followed me: a modified online Scrabble so a whole room could play together without seeing each other's boards, and a custom online Housie for the team.
Learned: Play pulls a group in faster than anything; the live-quiz product is built on the same instinct.
A review-and-catalog platform where friends who knew a domain (phones, courses) curated and wrote about it. Two developers, many contributors; ran for about two years.
Learned: Platforms outlive their traffic: one contributor parlayed his infoporch work into a job, then his own consultancy.
A run of viral Facebook apps that pulled friend data through the Graph API and composited it into shareable images: your top 6 friends cast as F.R.I.E.N.D.S characters, interaction leaderboards, many themes. People loved sharing them.
Learned: First taste of virality: build something people want to share and distribution takes care of itself.