Projects & Open Source

Open Source: SREGym Fault Scenario Contribution

2026 | Merged PR #886

🔧 Merged into SREGym (AI SRE Benchmark)

Contributed a fault scenario to SREGym, an open-source benchmark for evaluating AI agents on realistic system failures.

• Reproduced a real Kubernetes incident: modeled a documented admission-control failure (issue #128162) where multiple mutating admission webhooks’ cumulative timeouts exceed the API server’s global admission deadline, silently blocking all pod creation with no crash.
• Realistic, agent-resistant design: used real HTTPS admission servers and ecosystem-named decoy webhooks (not dummy backends) so agents can’t dismiss the setup, plus a deliberate “near-miss” policy that looks corrective but fails.
• Reward-hack-resistant oracle: built a four-property mitigation oracle culminating in a fresh probe-pod admission test that rejects fixes masking the symptom (e.g., deleting all webhooks) without restoring the required control-plane path.
• Stack: Python, Kubernetes (admission webhooks, NetworkPolicy, RBAC), Docker/kind

gf: GlassFish Dev Workflow CLI

2026 | GitHub Repository

⚡ Edit-Deploy Cycles: 2 min → 5 s

A cross-platform GlassFish development CLI that replaces IDE-driven deployment workflows with terminal-first hot-swap and AI-assisted server management.

• JDWP Bytecode Hot-Swap: Built a JDI-based hot-swap client that injects modified bytecode into running JVMs with automatic fallback to full redeployment on structural class changes.
• JVM & Framework Reverse Engineering: Diagnosed GlassFish XHTML staleness by tracing Mojarra Facelets cache internals and identifying ProjectStage-dependent refresh behavior.
• JasperReports Hot Reloading: Bypassed JasperReports classloader caching by resolving report resources directly from exploded deployment paths.
• AI-Augmented Workflow: Integrated as a Claude Code skill (/gf) enabling AI-driven deployment, server lifecycle management, and hot-swap operations from natural-language commands.
• Stack: Bash, Java (JDI/JDWP API), GlassFish 8, rsync, Claude Code Skills

Token-Efficient LLM Query Router

2026 | AMD Hackathon Act II, Track 1 | GitHub Repository

⚙️ Precision-Gated Cascade: Zero-Token Tiers First

An agent that answers 8 task categories (factual Q&A, math, sentiment, summarization, NER, code debugging, logic, code generation) while minimizing paid API token spend under a hard accuracy gate.

• Precision-gated cascade: a zero-token regex router (64/64 on dev and adversarial paraphrases), then a non-generative ONNX sentiment classifier shipped only after measuring 100% precision, then a LoRA fine-tuned Qwen2.5-1.5B for local factual Q&A and NER, with a cloud model as accuracy backstop. Each tier answers only what it can prove it handles; everything uncertain escalates.
• Program-aided math: instead of having the model compute, it emits a symbolic expression evaluated by a sandboxed AST interpreter. Arithmetic errors disappear as a category and output tokens drop about 90%.
• Verification guards over self-assessment: model-reported confidence measured unreliable at this scale, so local answers pass deterministic checks instead. An entity-completeness check catches 99% of dropped entities in NER output.
• Runtime speed calibration: detects the grading host’s actual throughput and enables a local tier only when it can finish in time.
• Evidence-driven iteration: an offline LLM-as-judge harness mirroring the official rubric tested every change before a submission was spent, across 17 measured releases with root-cause analysis for a production timeout, an accuracy-gate failure, and a token regression from an over-broad regex. Immutable Docker tags per submission kept every graded build reproducible.
• Stack: Python, PyTorch, Unsloth/LoRA, ONNX Runtime, Docker, Fireworks API

KnowledgeRelay: AI Knowledge Transfer Agent

2025 | GitHub Repository

🏆 6th Place, DSI AI Agent Hackathon 2025

Built in a 24-hour onsite hackathon to solve a problem every team has: when someone leaves, the unwritten knowledge leaves with them. The agent captures that knowledge before departure, then answers new joiners’ questions from it.

• Adaptive knowledge capture: interactively prompts an outgoing member with context-aware questions generated from the ongoing conversation, so the knowledge base is built by interview rather than by hoping someone writes documentation.
• Document and code ingestion: extracts knowledge from uploaded PDFs, Word files, and source code, with separate chunking strategies for code and prose.
• Grounded retrieval: rewrites follow-up questions into self-contained queries using prior conversation context, and attributes every answer back to its source file.
• Hybrid local/cloud LLMs: runs against OpenAI APIs or locally hosted Ollama models.
• Stack: Python, FastAPI, LangChain, ChromaDB, OpenAI, Ollama, React

Open Source: PrimeFaces JPALazyDataModel Enhancement

2024 | View Pull Request #12865

🔧 Merged into Official Library

Contributed a feature enhancement to PrimeFaces, a widely used open-source JSF component library.

• Custom Filter Injection: Added support for injecting custom FilterMeta into JPALazyDataModel, enabling advanced filtering beyond the stock component behavior.
• Framework Integration: Extended the DataTable filtering pipeline while preserving compatibility with existing PrimeFaces query generation patterns.
• Stack: Java, Jakarta EE (JSF), PrimeFaces, JPA

Shopaholic: E-Commerce Marketplace

2022 | GitHub Repository

A full-stack e-commerce platform connecting customers, suppliers, and partner banks with inventory and order management capabilities.

• Built REST APIs for authentication, product management, and order workflows using Node.js and Express.js.
• Developed a React/Redux frontend with MongoDB-backed inventory and transaction management.
• Stack: JavaScript, Node.js, Express.js, MongoDB, React, Redux