Hi, my name is
Rajeev Ranjan Chaurasia
Software Engineer
I build agentic systems and the developer infrastructure they run on.
About
I build agentic systems and the developer infrastructure they run on. I’m a Software Engineering graduate student at San Jose State University and an NVIDIA Certified Professional in Agentic AI. This summer I built CI infrastructure for NVIDIA’s Omniverse simulation pipelines, and before that I spent two and a half years at Tekion Corp on enterprise accounting and payment systems.
I also contribute to RocketRide, an open-source AI pipeline engine with a C++ core, where I work on the agent loop, RAG grounding and MCP tooling. Most of what I build ships with an evaluation harness and a control group.
Experience
- Owned the design and delivery of a dependency-aware CI test-selection system in GitLab CI/CD for NVIDIA's Omniverse Mega robotics simulation platform, adding an LLM layer that proposes file-to-test relationships static analysis alone could not see.
- Validated the system two independent ways before it touched production: replayed 628 historical merge requests, then ran a controlled comparison with the system disabled and enabled that measured a 50.5% reduction in executed job time.
- Extended the source-impact analysis to a second internal team's codebase, Omniverse Kit, adding Clang-based C++ analysis alongside the existing Python analysis on an entirely different build architecture.
- Contributed to an org-level Omniverse initiative exposing internal capabilities as open-source, agent-compatible libraries, working on API and interface design as part of a multi-contributor effort.
Open Source
rocketride-server
RocketRide8,421 stars4 merged · 1 in review
An open-source AI pipeline engine: a multithreaded C++17 runtime executing portable JSON pipelines across roughly 130 Python nodes, 15+ LLM providers, 9 vector databases, and four agent frameworks.
MergedDeliver each lane payload exactly once#2063
Swept seven handlers across five nodes that silently doubled their output, tracing the cause to the C++ engine's default-forward rule rather than the Python nodes themselves. Tests assert delivery counts instead of "the node forwarded", because the latter stayed true throughout the bug; the reviewing maintainer mutation-tested each node and called it "careful work, well evidenced".
Read the pull requestMergedDeliver each flushed embedding batch exactly once#2062
Found why every 64-chunk batch was delivered twice, and put the suppression in the lane handler rather than the flush helper, because close() routes through the same helper and would have left every downstream node open. Reverting the fix delivers 128 documents where the source sent 64.
Read the pull requestIn reviewGround chat answers in what retrieval returned#2084
Closed a RAG hallucination gap on both sides of the LLM call. Fabrication measured across five models dropped from 38 of 43 to 0 of 43, with a control group confirming that answerable questions were never suppressed and a separate check that JSON-mode answers still parsed.
Read the pull requestMergedMake agent list summaries answerable and repeats visible#2072
Diagnosed a production agent loop that ran 26 waves and roughly 400,000 tokens without an answer, because the planner only ever saw two rows of each tool result. Traced it to missing signal and permission rather than a missing safeguard, then measured the fix across models from several vendors.
Read the pull requestMergedAdd MCP authoring tools for agents building nodes#2082
Extended the in-engine MCP server with resolve_config and scaffold_node, and closed a validate_pipeline gap where a pipeline naming a provider that does not exist validated clean. The scaffolder is checked against the engine's own contract test suite rather than by inspection.
Read the pull request
Projects
Build Sleuth
Featured- Reads a failing GitHub Actions run, classifies the failure, ranks the culprit files, and opens a draft pull request only after the patch applies and the original job passes on a rerun inside a container.
- Measured 0.915 accuracy over 71 hand-labeled cases against a 0.218 regex baseline, on an evaluation harness whose negative control reruns every maintainer fix alongside a deliberately corrupted copy.
Cross Foot
Featured- Routes each document to deterministic extractors or a vision LLM by signature, reaching 96.02% auto-accept precision at 6.38% human review across 1,880 held-out fields.
- Multi-provider LLM infrastructure with token-bucket rate limiting and capability-filtered spillover, backed by an append-only cost ledger that recovered 97.1% of injected discrepancy dollars end to end.
leakproof
Featured- Finds credentials across full git history and decides which are real offline, with no call to any provider: CRC32 and key-parse proofs where the format allows one, and a published confidence table where it does not.
- 100% precision and recall over 145 credentials planted in source, from a generated corpus of 295 plants and 30 decoys, with fixture, vendored and documentation paths graded separately.
- Keeps its own worst regression in the commit log rather than tidying it away: an earlier build treated “this JWT decodes” as “this JWT is live”, which made 32 of 33 confirmed findings wrong.
- Skips the pytest tests a pull request cannot affect, and proves each skip with a witness that gets re-verified before pytest acts on it.
- Fail-closed CDC pipeline: PostgreSQL to Debezium to Kafka to Spark Structured Streaming to Parquet, then dbt into DuckDB and Snowflake. Publication is gated by row count, per-column checksum, aggregate, and delete reconciliation against source history; both warehouses must independently agree or the last validated dataset stays published.
- Distributed job orchestrator in Go: fenced leases with monotonic generations, DAG dependencies, deficit round-robin fairness, and a SQLite control plane whose scheduler decisions replay byte for byte, with a chaos and reconciliation harness that voids its own runs rather than publishing an ambiguous number.
- Co-built a multi-agent clinical platform with LangGraph, routing triage, symptom and summarization agents across a self-hosted medical model and two general-purpose models.
- Grounded medication answers in FDA DailyMed source text through a pgvector retrieval layer with a patient-medication metadata prefilter and a similarity floor, so the system answers from real source material or declines to answer at all.
Skills
Languages8
- Python
- TypeScript
- C++
- Go
- Java
- Ruby
- JavaScript
- SQL
AI & LLM Engineering22
- MCP (Model Context Protocol)
- AI Agents
- AI Agent Architecture
- Multi-Agent Orchestration
- Agent Evaluation & Tuning
- Retrieval-Augmented Generation
- Vector Databases
- LangGraph
- CrewAI
- Vision LLM Routing
- Prompt Engineering
- LLM Integration
- LLM Evaluation
- vLLM
- TensorRT-LLM
- SGLang
- Triton & CUDA Kernels
- Machine Learning
- Vertex AI
- Gemini
- NLU Concepts
- AI-Assisted Development
Testing & Verification8
- Mutation Testing
- Negative-Control Testing
- Contract Testing
- Evaluation Harness Design
- Benchmarking
- Chaos Testing
- Root Cause Analysis
- Query Optimization
Developer Infrastructure & CI/CD13
- GitLab CI/CD
- GitHub Actions
- Docker
- Kubernetes
- Helm
- Static Analysis
- Dependency Graph Analysis
- Test Impact Analysis
- Pipeline Optimization
- Policy Automation
- OpenTelemetry
- Prometheus
- Application Security
Backend & Distributed Systems17
- FastAPI
- Spring Boot
- Flask
- REST APIs
- gRPC
- Microservices Architecture
- Event-Driven Architecture
- Kafka
- RabbitMQ
- Stream Processing
- Apache Flink
- Apache Spark
- Change Data Capture
- Concurrency & Thread Safety
- Distributed Consensus (Raft)
- Spring Data JPA
- Hibernate
Cloud & Databases14
- AWS
- Google Cloud
- PostgreSQL
- pgvector
- Supabase
- MySQL
- SQLite
- Redis
- ClickHouse
- DuckDB
- MongoDB
- Elasticsearch
- dbt
- Apache Iceberg
Full-Stack & Integrations10
- React.js
- Next.js
- Tailwind CSS
- WebSockets
- Server-Sent Events
- Swagger/OpenAPI
- Chrome Extensions
- Google Maps API
- Telegram Bot API
- Serper API
Foundations & Practices9
- Data Structures & Algorithms
- Object Oriented Programming
- System Design
- Design Patterns
- Git
- Code Review
- Technical Writing
- Agile & JIRA
- Postman
Education & Credentials
Coursework
Certifications & Awards
NVIDIA Certified Professional: Agentic AI
NVIDIAIssued Aug 2026 · Expires Aug 2028
Validates designing, building, evaluating and operating agentic systems: agent architecture, cognition and memory, evaluation and tuning, data integration, human oversight, deployment at scale, and ongoing operations.
ClaraCare
DeveloperWeek 2026 HackathonFeb 2026
Winner of the Deepgram “Voice Operator” autonomous-agent prize. An AI companion that calls seniors daily, detects cognitive changes, and surfaces them to families through a real-time wellness dashboard.
SJ Hopes
SJHacks 2025Apr 2025
Winner of the Best for Homelessness track. Helps caseworkers find open shelter beds on a map, register clients, reserve beds and assign paid micro-tasks, with an analytics dashboard on top.
ClearClause
Frostbyte HackathonJan 2026
Winner at the Frostbyte Hackathon. An AI legal-document analyser that reads any contract and returns clause-by-clause risk analysis, fairness scoring, PII protection and voice-interactive Q&A in plain English.
Contact
I'm currently open to new opportunities. Whether you have a question, a role in mind, or just want to say hi, my inbox is always open.
Prefer email? rajeevchaurasia.dev@gmail.com