Problem: Processing large PDFs for semantic search was slow and inaccurate.
Solution: Async RAG pipeline with sliding window chunking and in-database vector search.
Result: 3min → 17s processing · 9/10 accuracy · 28% accuracy gain over fixed chunking · <5ms search latency

Problem: Navigating large codebases is slow and context is lost.
Solution: Hybrid search combining semantic embeddings with BM25 + AST-based chunking.
Result: 85% retrieval accuracy · 38ms latency · 2,000+ code chunks indexed

Problem: Manual PR reviews are slow and miss security vulnerabilities.
Solution: Event-driven agent processing GitHub webhooks with Groq LPU inference.
Result: Sub-200ms end-to-end · automated inline PR comments · real-time bug and vulnerability detection

Problem: Understanding Transformer internals beyond API calls.
Solution: 124M parameter decoder-only Transformer built in raw PyTorch, no shortcuts.
Result: 41% validation loss reduction (8.56→5.05) · gradient accumulation on dual T4 GPUs