AdvaitShinde

Image retrieval and captioning for failure analysis, served from on-premises GPUs.

• 92% Recall@1 from a CLIP model fine-tuned to retrieve past failure cases, saving 2–3 hours per turnaround.
• 10× throughput after replacing the Matplotlib image pipeline with CUDA rasterization (20 → 200 images/s), with a 2% better retrieval score.
• 84% F1 on generated failure descriptions (70% exact match): CLIP image embeddings pass through a Q-Former into a small language model, trained with LoRA.
• Under 50 ms per query in deployment, using thread-based parallelism.

LLM infrastructure for fintech clients: routing, inference cost, guardrails and retrieval.

• 10+ internal apps adopted the OWASP-aligned guardrails framework I built for bias, toxicity, PII and prompt injection.
• 20% lower inference cost from KV caching and speculative decoding in PyTorch.
• 50% faster contract generation for a client, through an agentic RAG workflow I co-designed.
• Benchmarked small language models (Llama, Qwen, Gemma, Mistral, Phi) on fintech NLP tasks to design an API gateway that routes between them on private GPU clusters.
• Delivered a multimodal RAG proof of concept on on-premises hardware with vision-language models, CLIP and ColPali embeddings.