LLM workflow engine
Designed and built the engine that composed multiple model types, including large language models, into configurable pipelines, before the commercial LLM boom made that a common product.
A selection across AI systems, trading and fintech infrastructure, and open source. Much of my client work is under NDA, so these describe the engineering rather than the client.
Designed and built the engine that composed multiple model types, including large language models, into configurable pipelines, before the commercial LLM boom made that a common product.
Agent workflows, retrieval-augmented generation, tool use and evaluation pipelines for client products, from first prototype through the checks that make them dependable in production.
AI agents for email. Forward any message to a dedicated address and an agent identifies the ask, plans an approach, uses tools such as web search, and replies. Processing is private and transient, and the core is open source and self-hostable.
A data product that compares 1,700+ SEBI-registered portfolio management schemes against 20+ indices across timeframes, turning public disclosures into evidence people can decide with.
A practitioner's course on reinforcement learning, from the fundamentals to working agents, published by Packt and listed on O'Reilly.
Order management on a persistent PostgreSQL ledger, all-or-none day orders so fill pairs stay symmetric, reconciliation after gap-up and gap-down opens, and crash and reconnect recovery without duplicate or orphaned orders. Paper-trade mode and kill switches before anything runs live.
A harness that replays historical data with explicit fill, slippage and latency assumptions, runs parameter searches without overfitting to one window, and reports the gap between backtest and live behaviour.
Led development of automated trading systems and strategy infrastructure: execution and order management, exchange connectivity, backtesting and deployment tooling, and production monitoring.
Off-chain opportunity detection with on-chain execution through smart contracts, submitted via private relays to avoid front-running and failed-transaction gas, plus a clear-eyed view of where such systems are viable.
Engineered services on the payments stack at one of India's largest e-commerce platforms.
Exploring and understanding Python through surprising snippets. Endorsed by Andrej Karpathy, featured on Hacker News and used as teaching material in university courses.
A blockchain built from scratch in Python with a step-by-step tutorial, originally published on IBM Developer.
A tool that inspects a repository and generates a configuration for coala, a language-independent static-analysis framework. I later returned to the programme as a mentor.
A knowledge-base product offered as SaaS and as a self-hosted deployment, co-founded and built as a side project.
More case studies at zobyt.com/work, and older experiments on my blog.