Kausshik Manojkumar
Software engineer who builds agentic systems with enforced safety boundaries - and the test infrastructure that makes them trustworthy. Software Engineer at RTX (Collins Aerospace), Cedar Rapids, IA. Iowa State University B.S. Computer Science, May 2025, GPA 3.97. ACM TKDD 2025 acknowledged contributor.
Current work
At RTX: one of four engineers who built black-box test infrastructure for a $600+ M avionics modernization program that had no way to test its software. It is now required for PRs, running in CI, referenced in official documentation, and used by 40 teams; 400+ engineers trained. Currently piloting agentic test-case generation that takes test cases from 16+ hours to roughly 30 minutes. Also 1 of 3 engineers defining migration strategy for the 2M+ LOC Ada-to-C++ modernization.
On the side: building an agentic assistant that cannot touch anything on its own. A typed proposal schema forms the trust boundary - the model inspects through read-only tools, emits only structured proposals from a fixed set of allowed action types, and an unknown action type fails deserialization, so it can never reach the write path. Every write is gated on human approval, and there is a redaction and data-handling layer in front of egress. Personal side work; no further details are public.
Contact
Projects
- QuillReader - Native macOS PDF reader with AI annotations, where the four-level privacy dial is enforced at the network boundary rather than promised in the settings copy. Crash-safe sidecar annotation persistence; never mutates the source PDF. Swift, SwiftUI, PDFKit, CryptoKit. Source.
- SearchAround - Non-linear mind-map search engine: queries become nodes on an infinite canvas and follow-ups branch from selected results. Prompt path hardened against injection from scraped web content - sources tagged as data, size-capped, strictly parsed, and failing open by keeping all sources. Next.js, reactflow, FastAPI, Exa. Live site.
- UniLibrary RAG - Retrieval system over hundreds of research papers from the Iowa State CS department. Hybrid retrieval (vector similarity + metadata filters) plus a validator model that catches hallucinated citations before display. Deployed to 3 university departments. Python, FastAPI, Qdrant, n8n, GCP. Source.
Skills
Languages: C++, C, Python, Swift, Java, TypeScript, SQL, Bash. AI/Agentic: agent loops and tool use, MCP, RAG pipelines, prompt-injection hardening, model evaluation, embeddings, fine-tuning, transformers. Backend/Systems: FastAPI, Flask, Spring Boot, REST, WebSockets, pub/sub and request-reply architectures, Linux, large legacy codebases. Infra/Testing: PyTest, CI, Docker, AWS, GCP, Qdrant, MySQL, DynamoDB, Git.
Publications
ACM TKDD 2025 - "Forecasting Graph-Based Time-Dependent Data with Graph Sequence Attention" (acknowledged contributor). Brought in mid-review to unblock validation: sourced and evaluated additional datasets and adapted a second model to run against them.
Machine-readable summary (llms.txt)