Saket Garodia

Problem: Developers need a code-building assistant that can clarify intent, plan architecture, and modify projects while keeping human approval in the loop.

Solution: Built a LangGraph-orchestrated 4-agent pipeline with Clarifier, Planner, Architect, and Coder agents plus sandboxed file, glob, regex search, and shell tools with dangerous-command blocking.

Problem: Consultation notes are often unstructured and slow to convert into reliable client-ready follow-ups.

Solution: Built a full-stack GenAI SaaS that converts notes into structured summaries, action items, and email drafts using real-time SSE streaming, structured LLM outputs, Clerk authentication, and subscription-gated access.

Problem: Users need to search and ask questions over a large Medium article corpus with both lexical precision and semantic recall.

Solution: Built a full-stack hybrid-search RAG system using BGE-base embeddings, BM25, reciprocal-rank fusion, FastAPI, Qdrant, and a Next.js frontend.

Problem: IT tickets need accurate classification while keeping deployment footprint practical.

Solution: Fine-tuned BERT-base on 47.8K IT tickets, then compressed the model via knowledge distillation into DistilBERT and FP16/INT4 NF4 quantization.

Problem: Large open-source models need practical alignment workflows that can run efficiently without full-model training.

Solution: Applied QLoRA-based SFT to Llama-2-7B, then aligned Mistral-7B with DPO on 12.8K preference pairs without a reward model.

Problem: Understanding LLMs deeply requires implementing core transformer components rather than only calling APIs.

Solution: Implemented a character-level GPT in PyTorch with multi-head self-attention, LayerNorm, residual connections, and a full training loop.