People · Business · Technology

Principal Engineer at Infrrd. I build AI systems and the teams that build them: enterprise document AI at production scale, and the engineering standards that keep it running.

• GenAI and machine learning architecture for a document extraction platform processing roughly 1M pages per day
• Agent tooling on the Model Context Protocol, and model routing and fallback across providers
• Fine-tuning small open models, evaluation harnesses, and treating cost per page as a first class metric
• Organization wide Python engineering standards across multiple teams

Generative AI and agents

Agent and LLM frameworks

Models, serving and routing

Training and data

Backend and architecture

Cloud and infrastructure

CI/CD and observability

Ways of working

Tool Attention Is All You Need Dynamic tool gating and lazy schema loading for scalable agentic workflows. arXiv:2604.21816 · repo

Locale-Conditioned Few-Shot Prompting for On-Device PII Substitution Mitigating demonstration regurgitation with small language models. arXiv:2605.13538 · repo

The Great Recalibration The 2026 pivot from generalist wrappers to application layers and industrialized AI services. SSRN:6071412

Three pending applications with the United States Patent and Trademark Office, on query optimization and entity extraction from documents: 19/375,298, 19/375,285, 19/432,186.

Tools and plugins

• research-anything — verified research pipeline: sources captured, claims bound to exact quotations, gated before anything ships
• content-repurposer-skill — Claude Code plugin for drafting technical content in your own voice
• openai-privacy-filter-benchmark — benchmark and inference harness for OpenAI's open-source PII detection model
• book-forge — build pipeline turning HTML into folio-stamped, self-audited PDF

On PyPI (all packages)

• excel-form-extractor — extract checkboxes and dropdowns from .xlsx and .xls files, zero runtime dependencies
• zenofai — the Zen of GenAI, printed on import
• quote-ondemand — fetch a quote from the internet

• anujsadani.in — site and profile
• tech.anujsadani.in — essays on AI engineering, infrastructure and security
• Dev Panda & Vibe Dragon — newsletter
• Medium

"When I wrote the code, only God and I knew how it worked. Now God alone knows!" — The eternal truth of legacy code

0.1 + 0.2 != 0.3 — It's caused by how they are stored in hardware. Read: Floating-point arithmetic

"The aim of the wise is not to secure pleasure, but to avoid pain." — Aristotle

"Architecting around entropy" — Managing complexity in distributed systems