A production-grade lightweight distributed task scheduling platform for small-to-medium teams. Benchmarked against XXL-Job, Kubernetes Job Scheduler, and Apache Airflow — designed to be leaner and easier to operate than all three.

Architecture highlights:

• Multi-instance Scheduler layer with Redis-based distributed lock for automatic leader election — zero single-point-of-failure
• DAG workflow engine for complex task dependency resolution
• Resource-aware scheduling: CPU/GPU/memory quotas enforced at dispatch time, preventing task starvation under load
• Multi-tenant resource isolation with per-tenant quota management
• Pluggable executor model: Shell / Python / Docker / HTTP

Engineering decisions:

• Optimistic locking + idempotency keys on all state transitions ensure exactly-once execution semantics
• Async processing pipeline with Redis task queue decouples scheduling decisions from executor availability
• Spring Cloud Gateway as API surface — rate limiting, circuit breaking, and distributed tracing from the start

An AI-powered astronomy knowledge platform backend. Combines RAG, tool-calling, and real-time streaming to deliver expert-level responses on undergraduate astronomy content.

Architecture highlights:

• RAG pipeline: knowledge vectorized and stored in Qdrant; retrieved at query time and injected into prompt context
• SSE (Server-Sent Events) streaming endpoint for real-time AI response delivery with sub-second first-token latency
• Tool-calling layer integrates Wikipedia search, vector knowledge base queries, and concept card lookups — model selects tools autonomously or via explicit user directives (@wiki:, @kb:, @card:)
• Configurable advisor chain: Memory + RAG + Wikipedia advisors compose in runtime, each independently toggleable

Engineering decisions:

• OpenAI-compatible API abstraction layer allows hot-swapping between DeepSeek, SiliconFlow, and other providers without code changes
• Per-tool disable flag on explicit invocations prevents duplicate calls when the user prefetches a tool result manually
• Session-scoped conversation history enables coherent multi-turn dialogue without client-side state management

An autonomous AI agent that conducts deep multi-source research and generates structured reports. Given a query, it plans, crawls, analyses, and writes — without human intervention at any step.

Architecture highlights:

• Four-stage pipeline: planner.py → crawler.py → analyzer.py → report.py — each stage is independently testable and replaceable
• Parallelized web crawling with Selenium headless Chrome; automatic quality filtering rejects low-signal pages before analysis
• Multi-modal analysis: text via o3-mini for summarization; screenshot capture + vision model for tables and figures that can't be parsed as text
• Outline-first report generation: LLM builds a hierarchical section structure with weighted priorities before writing any prose

Engineering decisions:

• Separating planning from execution (distinct planner.py / research_agent.py) keeps the agent loop auditable — every search decision is traceable back to an explicit plan step
• Source citation embedded throughout generation, not appended at the end — ensures claims remain grounded even during iterative rewrites

A real-time voice conversation system with speech-to-text, LLM inference, and text-to-speech in a single low-latency pipeline. Supports multiple languages with minimal perceptible delay.

Architecture highlights:

• WebSocket-based full-duplex audio stream between browser and Node.js backend — no HTTP polling, no round-trip overhead
• audioWorklet.js runs in the browser's audio thread for Voice Activity Detection (VAD), sending only speech segments — reduces upstream bandwidth and LLM prompt noise
• LLaMA inference integrated server-side; response text piped directly into TTS synthesis before the full reply is complete (streaming TTS)

Engineering decisions:

• Decoupling VAD into the AudioWorklet (off main thread) prevents UI jank during continuous recording
• Next.js frontend with Tailwind CSS keeps the interface latency-free; no heavy state management framework needed for a single-session voice app
• Real-time latency monitoring logged per-session — latency data used to tune VAD silence threshold and chunk sizes