Imaad Zaffar Khan

Software Engineer → AI Engineer

Taazaa Inc. · Ohio, United States

• Led AI development for a municipal tax platform (CiVi), designing and implementing the document-intelligence system used for tax filing and municipal review.
• Direct the technical design and ongoing development of an MCP gateway for a multi-tenant marketplace platform (Arcadier), defining how AI assistants interact with its existing APIs.
• As part of Taazaa's core AI research team, review current research and prototype ideas in agentic systems, retrieval architectures, transformer models, model evaluation, and responsible AI for use in client products.
• Provide technical oversight for junior engineers on AI projects, defining scope, reviewing designs, and ensuring code quality and delivery.

Research Assistant

University of Illinois Urbana-Champaign · Illinois, United States

• Developed continual-learning methods for the DARPA-backed ECOLE project, enabling computer-vision models to learn new object concepts without losing previously acquired knowledge.
• Designed experiments around catastrophic forgetting, concept prioritization, and region-based learning using DINOv2, Alpha-CLIP, and Semantic SAM.

Data Engineer

Larsen & Toubro Infotech · Maharashtra, India

• Built and maintained Talend-based ingestion and transformation pipelines for large enterprise datasets used in reporting and downstream analytics.
• Developed data-validation and reconciliation workflows that caught quality issues earlier and reduced repetitive manual checks for the operations team.

CiVi

Python · Kafka · AWS · Azure AI · Gemini

CiVi is a municipal income-tax platform. It puts paper filings, online submissions and auditor review in one place, and it's live in Dublin, Ohio.

I designed and built every AI component in it. The city sends bundles of several hundred scanned pages at a time, with federal, state and municipal forms mixed together in no particular order. The pipeline cuts a bundle into separate returns, works out which form each one is, reads the fields off the page, then hands clean data back to whatever system asked for it. Three paths feed it: bulk scans, the public e-filing portal and the internal auditor portal. Each city's data stays separate, so adding a new one is config, not code.

Extraction picks its tool by form type. Azure Document Intelligence ships prebuilt models for the common federal forms, trained on those exact documents, so it reads them accurately. Municipal and state forms have no prebuilt model, so those go to Gemini instead. Everything moves as Kafka messages rather than requests, because a six hundred page bundle takes tens of minutes and nothing that slow belongs in a synchronous call. Work that fails gets caught, retried with backoff, and can be run again by hand.

Taazaa has published its own account of the project, covering the business side: AI assistant cuts tax return processing time from days to hours.

Arcadier MCP Gateway

C# · ASP.NET Core · MCP · Redis

Arcadier is a multi-tenant B2B2C marketplace. Its MCP gateway lets people run most marketplace operations in plain language, through the REST APIs the platform already has.

I designed it as a vendor-neutral translation layer, so it works with more than one AI assistant and more than one marketplace without changing the platform underneath.

Multi-Repository Code Intelligence

Python · LangGraph · Neo4j / Memgraph · Qdrant

A code-intelligence platform you can talk to. Separate indexer, agent and MCP services answer questions across repositories and trace how services, files and dependencies connect.

I work on the indexing and on the analysis workflows for architecture, security and data access, pairing dependency graphs with vector search.

Source: repo-indexer, repo-agent, repo-mcp-server, auth-gateway

The file-shaped hole in MCP

Every AI product takes a file attachment. MCP, the protocol built to connect those products to real systems, has no file input type at all. The reason isn't size. It's that by default the model has to type the bytes out itself, which puts a 126 KB screenshot at around 112,000 output tokens. Five workarounds ranked by how far each one survives, the upload-ticket pattern I ended up with after building it three times, and why the official draft is landing in the same place.

ECOLE

Environment-driven Conceptual Learning · Research Assistant, University of Illinois Urbana-Champaign, in partnership with UCLA

A program funded by the Defense Advanced Research Projects Agency (DARPA), on machine perception systems that keep learning after they ship. The program target was recognizing more than 1,000 object concepts in real time. My work was on concept scoring and on the wake-sleep training loop that decides what the model spends its time learning without losing what it already knew.

Enhancing Road Safety: Real-Time Detection of Driver Distraction through Convolutional Neural Networks

Preprint · Amaan Aijaz Sheikh, Imaad Zaffar Khan · arXiv:2405.17788

Compares CNN architectures for spotting driver distraction in real time, testing simple baselines against batchwise and non-batchwise fine-tuning of VGG16 and VGG19, plus a custom architecture with a transformer.

Graph Neural Network and NER-Based Text Summarization

Preprint · Imaad Zaffar Khan, Amaan Aijaz Sheikh, Utkarsh Sinha · arXiv:2402.05126

Extractive summarization built on graph neural networks and named entity recognition. The GNN models how text segments relate to each other, and NER keeps the summary anchored to the entities that carry the meaning.

Master of Computer Science

University of Illinois Urbana-Champaign · GPA 3.95 / 4.00

B.Tech., Computer Science Engineering

SRM University · GPA 3.9 / 4.0