# Mayank Saini

> AI engineer in India, building systems around agents. Public profile, updated 19 October 2026.

## Agent orchestration and financial-close systems

Mayank is an AI engineer working on LangGraph agent orchestration, MCP tool integrations, context management and financial-close data platforms. His machine-learning research studies routing among specialised models.

At Mirian AI, August 2026 to present, he is a Member of Technical Staff, AI Engineer. He owns FinClose end to end, from financial domain modelling through production operation, and designed and built the canonical operating data layer consumed by Mirian agents. His high-level work includes ERP and bank feeds, common source data, scoped access, source lineage and reviewable agent actions. Client data and implementation details are private.

At PwC US AC R&D, July 2024 to August 2026, he worked as a Machine Learning Engineer and was promoted to Senior AI Engineer. PwC One brought 50+ AI tools from Advisory, Tax and Assurance into one agentic layer. His work includes a LangGraph supervisor with MCP and A2A, context pruning/compaction/summarisation, Redis checkpointing, LiteLLM caching and Langfuse tracing. This portfolio describes his contribution; it does not reproduce an employer platform.

[Explore the enterprise agent case study](https://mayanksaini.dev/#enterprise-agents) and [financial-close engineering work](https://mayanksaini.dev/#financial-close) on the portfolio.

During an internship at Samsung R&D, he worked on a Smart TV recommendation project.

## Research

Two public arXiv preprints:
- [One Supervisor, Many Modalities: Adaptive Tool Orchestration for Autonomous Queries](https://arxiv.org/abs/2603.11545), 2026.
- [Towards Resource-Efficient Multimodal Intelligence: Learned Routing among Specialized Expert Models](https://arxiv.org/abs/2511.06441), 2025.

The 2026 preprint reports 67% lower cost against its matched hierarchical baseline across 2,847 queries. That result is specific to the study and is not a universal production savings claim. Preprints are not presented as peer-reviewed publications.

## Relay prototype

[Relay](https://mayanksaini.dev/relay.md) is a personal prototype for sharing technical artifacts while preserving source, history and access rules. Mayank built the interactive product mockup and launch film and worked through the architecture. Backend implementation is underway; the complete service is not released.

## Agent Recall

[Agent Recall](https://github.com/itsMay21/agent-recall) is an MIT-licensed TypeScript/Node CLI for local transcript recall across coding agents. It compiles prior sessions into a shared, redacted JSONL index that the next agent can search and read, with no model in the compiler. The public [v0.2.0 release](https://github.com/itsMay21/agent-recall/releases/tag/v0.2.0) includes a synthetic demo and setup/limits documentation.

## Agent Expression

[Agent Expression](https://github.com/itsMay21/agent-expression) contains seven portable skills for clearer agent-written text and code.

## Field Notes reading desk

[Field Notes](https://mayanksaini.dev/reading/) is a personalized local fork of ThariqS/ai-newtab, focused on agents, context, systems and interface design. Its local edition combines bounded public-source acquisition with Codex-authored summaries and separate reading questions, retaining source dates and reading scope. The portfolio shows a separately dated sample of 14 public-source entries with deterministic ranking and reading prompts. Each sample entry has an original generated editorial illustration of a specific source idea, with a caption stating its reading scope. The illustrations use the corresponding article, release notes or preprint abstract; they are not publisher artwork or independent verification of a source's claims. Optional interactive diagrams provide another way to inspect the idea. The local edition does not access browsing history or signed-in pages.

## Education and toolkit

Electrical Engineering at the Indian Institute of Technology Kanpur, with a minor in Machine Learning.

His public work names LangGraph, LangChain, MCP, A2A, Python, FastAPI, Redis, FastMCP, LiteLLM, RAG, Langfuse, AWS, Azure and Git.

## Interests

Mayank likes side projects, reading agent research and exploring how interfaces make technical ideas easier to understand.

Two questions he keeps coming back to:
- How should search change when agents do the browsing? He is interested in evaluations that explain where search went wrong, source evidence and preventing retrieved instructions from hijacking an agent's task.
- What was here before “here” existed? A playful philosophical question about the origin of the universe, what “before” means and simulation ideas.

## Contact and public evidence

- [Portfolio](https://mayanksaini.dev/)
- [Current one-page CV](https://mayanksaini.dev/Mayank-Saini-CV.pdf)
- [GitHub](https://github.com/itsMay21)
- [LinkedIn](https://www.linkedin.com/in/mayanksaini-ai/)
- Professional email: itsmayank.saini21@gmail.com

For details beyond this public profile, contact Mayank.