I'm Mayank. I build agent systems and financial data platforms. I'm curious about what makes an agent useful beyond a single answer: the tools it can use, the context it keeps, and the decisions we can inspect.
IIT Kanpur PwC US R&D Mirian AI
From the orchestration layer
to the operational details.
A place to share technical artifacts with their source, version history and access rules intact.
Explore the prototypeI built the interactive product prototype and launch film, and designed the sharing, rendering and agent-access architecture.
Owning FinClose end to end, alongside the common data layer its agents work from.
Agents prepare, reconcile and explain close tasks. Scoped access, source lineage and reviewable actions make it possible to check a result before it changes a financial workflow.
ERP and bank feeds flow into a common operating layer, so each new agent can work from consistent source data instead of another custom integration.
High-level account of my work. Client data and implementation details are private.
Connecting 50+ AI tools from Advisory, Tax and Assurance in one agentic layer.
I worked on a LangGraph supervisor architecture with MCP and A2A, letting an orchestration layer delegate work to specialised tools.
Pruning, compaction and summarisation help manage long-running context. Redis checkpointing, LiteLLM caching and Langfuse tracing support state, cost and inspection.
An account of my contribution, not a public reproduction of an employer platform.
Our 2026 preprint reports 67% lower cost against its matched hierarchical baseline across 2,847 queries.
My research with Arit Kumar Bishwas studies learned routing among specialised models and supervisor-based tool orchestration across modalities.
The 2026 preprint reports a 67% cost reduction against its matched hierarchical baseline, evaluated across 2,847 queries. The reported result is specific to that study.
Read the methodology on arXivShowing all 4 projects.
Public research. Methods,
experiments and trade-offs.
Adaptive Tool Orchestration for Autonomous Queries
2025 · arXiv preprintLearned Routing among Specialized Expert Models
Co-authored with Arit Kumar Bishwas.
Follow a synthetic task through tools, context, routing and review. This is an authored illustration of the decisions around an agent.
Find the evidence in a document and an image. Explain the answer.
Pick a task and run it. Select any node to inspect its job.
Synthetic inputs. Browser simulation. No live models or employer systems.
How a task is split, which tool is allowed to act, what context survives, and when a person needs to review the result. That's where I spend my time.
A question about agents, tools or the work behind them? Search the project notes.
Local keyword search across these project notes. No chatbot or external API.
I'm interested in evals that explain where a search went wrong, evidence that survives the hand-off, and keeping instructions in a retrieved page from hijacking the task.
How the universe began, what “before” even means, and whether this is all a simulation. I have a few questions for whoever wrote the spec.
I like getting into the details between a promising model and a system people can use. Building the prototype, reading the research and testing the assumptions are all part of that for me.
I studied Electrical Engineering at IIT Kanpur, with a minor in Machine Learning. Before my work at PwC and Mirian AI, I built a Smart TV recommendation project during an internship at Samsung R&D.
Explore my GitHubTools matter when the system around them works.
Side projects, agent research
and interfaces worth thinking about.
I enjoy making small tools, reading about how agents work, and exploring how an interface can make a technical idea easier to understand. A few things I keep working on:
