Hey, I’m Isuru

Client-side audio demuxing and slide keyframe extraction via HTML5 Canvas & Web Workers reduces network upload payloads by 96% (from 4GB down to ~20MB).

High-concurrency AssemblyAI Universal-2 engine with dynamic key rotation pool, speaker diarization, and word-level timestamp alignment.

Google Gemini 2.5 Flash curriculum planning pass followed by holistic synthesis merging audio transcripts with client-side OCR keyframes.

Client-rendered vector diagrams (Mermaid.js), mathematical proofs (KaTeX), statistical charts (Vega-Lite), and slide hotspot overlays at 99.9% lower cost than raster generation.

Decoupled Cloudflare R2 asynchronous job buffering surviving client disconnects, paired with a persona-driven Socratic RAG conversational tutor.

AST markdown parsing with mistletoe, converting formulas to native Word MathML (.docx), publication-ready PDFs, and interactive GFM.

High-throughput Kafka streaming architecture with real-time event validation and automated feature store ingestion.

LightGBM and gradient-boosted decision trees trained with continuous online feature engineering and hyperparameter tuning.

Ultra-low latency FastAPI microservices with model quantization, Redis caching, and automated drift detection.

Real-time WebSocket dashboard displaying live probabilistic distributions, SHAP feature impact, and confidence intervals.

Stateful LangGraph orchestration with recursive planning, dynamic tool selection, and automated error recovery.

Dual-tier memory architecture uniting short-term conversational context with long-term pgvector semantic knowledge retrieval.

Rigorous schema validation and safety guardrails enforcing deterministic JSON outputs and permission boundaries.

Distraction-free Next.js interface with live thought-stream visualization, tool execution inspect panel, and step debugging.