Projects
End-to-end AI products — from research prototypes to production deployments. Every project measured with real numbers, not just demos.
GraphRAG Module
Knowledge graph + RAG module as part of the company's broader RAG system. Proposed and led the project from concept to production-ready code.
🔍 Reranking Evaluation
Systematic evaluation across two datasets. Landed on configuration that improved metrics by +7-11 percentage points over fixed baseline, while keeping per-query latency under 2 seconds — the deciding factor for production recommendation. Integrated into production.
🐛 Critical Bug Resolution
Traced and resolved a significant bug in chunk retrieval during entity-filtering work. Once fixed, recall improved by roughly 20-80% depending on settings, across multiple datasets.
🧬 Ontology Extraction
Experiment-driven product improvement through systematic ontology extraction from corpora. Research doesn't always resolve in one pass — the iteration itself produced the improvement.
⚡ Production Status
System is production-code-ready. For complex corpora it requires meaningful compute, so tuning has been part of the work (postponed due to current focus on research project), but considered a delivered, completed piece.
SceneFlow Video Agent
Co-founded startup project — handled the entire technical stack while participating in product positioning, fundraising, market research, and beta testing guidance. Video agent with comprehensive tool integration.
🤖 Agent Engine
7000+ line engine with L0/L1/L2 routing. L0 list/get skips the LLM entirely — 200 L0 + 200 L2 classify_turn calls stayed under 2000ms wall budget.
🎨 Media Generation
Integrated Seedream (image) and Seedance (video) generation. Persisted n=1 Ark Seedream image and n=1 Seedance 5s 720p video with downloadable outputs.
💰 Cost Tracking
DeepSeek cloud cost_sum 6.16e-06 USD (n=2 synthetic). Bedrock Nova Micro cost_sum 1.05e-06 USD. Ollama stays null, not 0.0 — honest accounting.
🔌 MCP & Tools
Read-only MCP retrieve over studio HTTP. SceneFlow MCP limited to list/get only — security-first design with tool-use capabilities.
IntelliScribe
Knowledge-ingest + writing application. Full-stack with FastAPI backend, Next.js frontend, and Docker Compose deployment. Chat and generate share SSE streaming.
🔄 RAG Migration
Replaced in-repo Chroma with studio HTTP retrieve. Measured p95 228.0ms at n=50, 7.67 rps, warm 8-doc index, no LLM on the path.
📊 Retrieval Quality
20-question writing eval scored recall@5 = 1.0 (20/20) on fixture overlap retriever. That number is the question-set baseline, not live Chroma quality.
⚡ SSE Streaming
OpenAI-compatible path uses stream=true and yields SSE deltas. 13 tests green. Stub still slices 12 chars on purpose for offline tests.
🏷️ AI Disclosure
Disclosed AI generation as JSON boolean ai_generated on studio and product /health endpoints.
ai-studio RAG Services
Modular RAG service layer providing retrieval, evaluation, and cost tracking to multiple AI products. Designed for scalability and extensibility — each component can be independently deployed and measured.
🧩 Modular Architecture
Separate modules for retrieval, evaluation, cost tracking, and health monitoring. Each service can be consumed independently by different products, enabling flexible composition.
📈 Eval Framework
Systematic evaluation with measured quality metrics. Recall@5 0.933, MRR 0.778, p50 194ms. Honest about what each number represents.
💰 Cost Tracking
Per-query cost calculation from public token pricing. DeepSeek 6.16e-06 USD, Bedrock Nova 1.05e-06 USD. Ollama stays null, not 0.0 — transparent accounting.
🔌 Service Interface
HTTP API for retrieval, OpenAI-compatible streaming, health endpoints with AI disclosure. Products consume services, not implementations — enabling backend changes without breaking clients.
NLQxform — KG Question Answering
Master's thesis research on knowledge graph question answering. Built a language model-based question to SPARQL transformer, then extended it into an interactive scholarly QA system.