B.Tech Information Technology @ Manipal University Jaipur
I train models at the intersection of vision and language and I write about it too.
🩻 Medical Image Captioning — Frozen ViT-Base → projection MLP → Phi-2 (2.7B) with LoRA, trained in 5.5 hours on one RTX 5090. Evaluated on 500 ROCOv2 images:
| BLEU-1 | BLEU-2 | BLEU-3 | BLEU-4 | METEOR | ROUGE-L | CIDEr |
|---|---|---|---|---|---|---|
| 0.1362 | 0.0733 | 0.0361 | 0.0187 | 0.0636 | 0.1638 | 0.1045 |
Beats DS@BioMed on every metric, edges Kaprov on METEOR, trails CS_Morgan. Written up for Springer LNCS as an efficiency result: one consumer GPU instead of a cluster. Also documents a mode-collapse failure mode, caught with a zeros-gap test that checks whether the model is really using the image.
🧭 Failure Direction — Speech enhancement fails two opposite ways: deletes the words, or leaves the noise. I built a reference-free index that tells the two apart and now use it to predict, from the noisy input alone, which failure a model will hit even on unseen noise. Currently at the go/no-go gate before the prediction stage.
🧠 GiWaLM — 125M-parameter language model, pretrained from scratch on ~1B tokens with a custom 32K BPE tokenizer.
Ex-AI Intern, Wipro — Data Analytics & AI team, working across the GCP stack: Vertex AI, BigQuery, and scalable ML pipelines. Picked up Time Series, RAG, and LLMOps along the way, plus a stack of Anthropic Claude certifications (Claude Code 101, MCP, Agent Skills, Subagents).
💻 Languages
🧠 AI • Deep Learning • NLP
☁️ Cloud & MLOps
⚙️ Tools & Automation
🎨 Frontend
Where I spend my time: multimodal LLMs, parameter-efficient fine-tuning, AI automation, computer vision, NLP.
Anthropic certifications earned during my Wipro internship, alongside coursework in Time Series, RAG, and LLMOps.
Research collaborations, technical internships, and AI engineering roles in deep learning and multimodal AI.