I build multi-agent systems and the retrieval that keeps them honest. AI Engineer at Boston University's Questrom Computational Lab and Lead Backend Engineer at the Questrom Center for Action Learning. M.S. in Computer Science, graduating December 2026. NVIDIA-Certified Professional: Agentic AI and Claude Certified Developer. First author on an IEEE paper in automated program repair. Available January 2027 (F-1 OPT, STEM OPT eligible).

_{features · experience · open source · selected work · publications · certifications · stack · where this breaks · contact}

Features

• Agents in production: a 14-tool LangGraph agent on Azure OpenAI for the enterprise client Sikich.
• Measured retrieval: hybrid BM25 + vector search with Cohere re-ranking; precision@1 0.86, MRR 0.93.
• Evals as a release gate: 16 agent evals run 5 times before every portfolio release.
• Agent infrastructure: a Claude Code harness with least-privilege permissions, context-routing hooks, and CodeRabbit review linked to Linear through MCP.
• Guardrails: PII and entity redaction before retrieval; cited answers streamed over FastAPI SSE.
• Upstream: open PRs to vLLM and LlamaIndex; a Deep Agents bug fixed by a maintainer from my report.

Add-ons: NVIDIA-Certified Professional: Agentic AI · Claude Certified Developer

Boston University, Questrom Center for Action Learning · Lead Backend Engineer, August 2026 to present An agentic SDLC harness that holds Claude Code agents to the same gates as human engineers in a regulated client codebase: least-privilege permissions, context-routing hooks, CodeRabbit review linked to Linear through MCP, shipped as a plugin. A Django REST Framework and PostgreSQL backend with data-layer RBAC, revocable JWT, audit logs and an idempotent recurrence engine; CI fails when the API contract drifts from the React frontend.

Boston University, Questrom Computational Lab · AI Engineer, Graduate Researcher, May 2026 to present A production agentic RAG platform on Azure for the enterprise consulting client Sikich, 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. Customer-facing requirements, demos and weekly knowledge transfer.

Boston University · Graduate Teaching Assistant, MET CS 664 Artificial Intelligence (about 40 graduate students), September 2026 to present

IMG Systems Pvt. Ltd. · 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 Pvt. Ltd. · Associate Software Engineer Intern, Remote, 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.

Status and star counts below are rewritten weekly from the GitHub API by a workflow in this repository, so a merge shows up here without anyone editing the page.

Project	Contribution	Status	Stars
vLLM	PR #56846: pythonic tool parsers agree across streaming and non-streaming modes, from my report #56840	PR, under review	93.2k ★
LlamaIndex	PR #23053: run-level max_iterations and early_stopping_method apply on a reused Context, from my report #23051	PR, closed	52.4k ★
LangChain Deep Agents	issue #4846: CompositeBackend.ls("/") swallowed default-backend errors; reproduction filed; maintainer fix #4925 credits my report	issue, fixed	29.9k ★
MCP Python SDK	issue #3490: a rejected connect_to_server leaves its transport running	issue, open	24.5k ★

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.

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.

Agentic portfolio · source A LangGraph agent that answers questions about my work from a grounded fact file: BM25 fused with vector retrieval, answers streamed over SSE, precision@1 0.86, MRR 0.93, 1.16 s median. A facts gate fails the build when a corrected fact reappears, and 16 agent evals run 5 times before a release. Providers fail over in order, so an unset key degrades instead of erroring.

flowchart LR
  Q([question]) --> R[retrieve: BM25 + vector fusion]
  R --> A[LangGraph agent]
  A --> S([cited answer over SSE])
  F[(grounded fact file)] --> R
  E{{16 evals x 5 runs + facts gate}} -. blocks release .-> A

mindmap
  root((Kushal))
    Agents
      LangGraph
      14-tool enterprise agent
      MCP
      Claude Code harness
    Retrieval
      BM25 and vector fusion
      Cohere re-rank
      pgvector
    Evaluation
      LLM-as-a-judge
      16 evals x 5 runs
      P@1 0.86 and MRR 0.93
    Backend
      FastAPI
      Django REST Framework
      PostgreSQL and Redis
    Cloud
      Azure OpenAI and AI Search
      Docker and CI/CD
      AWS and GCP
    Upstream
      vLLM PR 56846
      LlamaIndex PR 23053
      Deep Agents issue 4846

Deep Learning for Contextual Bug Detection and Automated Fixes in Software Systems ICAICCIT 2024, IEEE, pp. 624-629. First author. Transformer plus GNN: 91.4% accuracy vs 88.2% for the transformer alone. IEEE Xplore · 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.

NVIDIA-Certified Professional: Agentic AI (NCP-AAI, 2026) is a proctored vendor exam, verifiable on Credly. So is Claude Certified Developer - Foundations (Anthropic, 2026). Plus AWS Cloud Technical Essentials (2026), Google Cloud Fundamentals (2025), and three completed courses of IBM's RAG and Agentic AI programme, each individually verifiable.

Written because a profile that only lists strengths is not worth reading, and because these are the questions an interviewer asks anyway.

kushal7887pd@gmail.com · LinkedIn · portfolio · Boston, MA

If you are working on multi-agent coordination or agent evaluation, I would rather compare notes than network.

_{back to top}