I build multi-agent systems and the retrieval that keeps them honest.
AI Engineer at Boston University's Questrom Computational Lab, MSCS candidate, graduating December 2026. First author on an IEEE paper in automated program repair. Available January 2027.
Portfolio · LinkedIn · ORCID 0009-0009-9318-1616 · kushal7887pd@gmail.com
- null in Kushal9889/BU-Life-AI
- null in Kushal9889/CF_Ladder
- null in Kushal9889/Cyber-Physical-Systems-and-the-Future-of-Urban-Living-Decision-Making-Challenges-and-Opportunities
- null in Kushal9889/Deep-Learning-for-Contextual-Bug-Detection-and-Automated-Fixes-in-Software-Systems
- null in Kushal9889/Intelligent_document-QA-system
BU Life AI · source A campus assistant for Boston University students, live with real traffic. A LangGraph supervisor classifies intent and routes to one of 3 specialised ReAct agents, each owning its own thread so concurrent users never share state. Retrieval fuses BM25 with NV-Embed 1024-dimension vectors over pgvector through an EnsembleRetriever. That routing decision is what cut redundant LLM calls by 70%.
The trade-off worth asking about: orchestration complexity bought state isolation. One agent with a long prompt was simpler and mixed tool namespaces across housing, dining, and events until retrieval started contaminating.
Contextual bug detection · IEEE ICAICCIT 2024, first author A transformer reads tokens and syntax; a graph network reads module dependencies; the two are concatenated and scored together. The combined model reached 91.4% accuracy against 88.2% for the transformer alone. The graph branch scores lowest on its own at 85.7%, which is the point: structure without content cannot tell a correct function from a broken one.
Enterprise document intelligence · Boston University, Questrom Computational Lab A production agentic RAG platform on Azure for an enterprise consulting client, owned from ingestion through deployment. A LangGraph agent exposing 14 tools for document question answering, comparison, and template-driven generation. Hybrid retrieval with LLM query rewriting and Cohere re-ranking, LLM-as-a-Judge evaluation for hallucination rate, PII redaction guardrails, and a Cosmos DB Gremlin knowledge graph linking clients, projects, and technologies.
Reported langchain-ai/deepagents#4846: CompositeBackend.ls("/") aggregated results at the root and discarded errors from the default backend, so a caller whose backend had failed saw a healthy but nearly empty filesystem. Filed with a reproduction and a proposed fix mirroring the existing grep root-merge check. A LangChain maintainer authored and merged the fix in #4925 three days later, crediting the report by name.
I did not write the patch. deepagents restricts merges to organisation contributors. What the report demonstrates is the part that transfers: reading an unfamiliar production SDK closely enough to find where it contradicts its own documented invariant, and writing it up precisely enough that someone senior acted without needing to ask a question.
Deep Learning for Contextual Bug Detection and Automated Fixes in Software Systems ICAICCIT 2024, IEEE, pp. 624–629. First author. IEEE Xplore · doi:10.1109/ICAICCIT64383.2024.10912101 · repository
Cyber-Physical Systems and the Future of Urban Living IGI Global, 2024. Co-author. Repository
Both repositories carry a CITATION.cff, so GitHub's Cite this repository button returns the correct BibTeX rather than a citation for the code.
Written because a profile that only lists strengths is not worth reading, and because these are the questions an interviewer asks anyway.
BU Life AI does not survive its own success. It runs on a Render free tier. The first thing to fail under load is CPU throttling and cold starts, then Neon connection limits. The fix is a paid tier with persistent workers and PgBouncer pooling. I have not needed it and have not pretended otherwise.
The paper's method needs data most teams do not have. It depends on a large corpus of code annotated with bugs and their fixes, plus runtime metadata. Where that corpus is thin, accuracy degrades. Generalisation across languages is untested, and the paper says so.
My first Pydantic schemas at IMG Systems were too strict. Documents that were merely unusual got rejected alongside genuinely malformed ones. I fixed it with fallback validators and logging on the rejection path, which turned silent data loss into a visible signal. That is the mistake I would tell you about unprompted.
Boston University, Questrom Computational Lab, AI Engineer, Graduate Researcher. May 2026 to present.
IMG Systems, Software Engineering Intern, Remote. August 2024 to April 2025. Extended a Python document-parsing pipeline on Apache Tika, raising extraction accuracy 20% across more than 5,000 candidate profiles a month and cutting recruiter screening time 15%. Pydantic structured-output validation against a JSON Schema reached 95% schema accuracy. Containerised services on PostgreSQL and Redis with Docker trimmed REST latency 25%.
Growaza, Associate Software Engineer Intern, India. January to July 2024. Cut API response time 30% with in-memory caching and asynchronous request handling, lifting engagement 22% for more than 1,000 daily active users. MySQL inventory dashboard tracking 2,000+ SKUs. JWT and role-based access control on AWS EC2 and S3.
NVIDIA-Certified Professional: Agentic AI is a proctored vendor exam, verifiable on Credly. Plus AWS Cloud Technical Essentials, Google Cloud Fundamentals, and three completed courses of IBM's RAG and Agentic AI programme, each individually verifiable.
If you are working on multi-agent coordination or agent evaluation, I would rather compare notes than network.