"I turn unreliable LLM outputs into production systems."
"AI reliability is a design problem, not a model problem."

• 96.1% mAP50 / 88.9% mAP50-95 on dense document layout and figure extraction via fine-tuned DocLayout-YOLO (YOLOv10m).
• ~95% cost reduction (to ₹0 API cost) by replacing raw LLM coordinate guessing with an LLM-to-IR 16-tool deterministic geometry compiler on local dual RTX 3090s.
• 13,000+ students served by campus digital platforms as Institute Web Convener at IIT Bombay.
• 3.4M+ transactions & orders analyzed in production retail basket analytics.

Production RAG Platform on Azure Container Apps · Live App

• Converts full playlist-scale lecture series into structured chapter notes, interactive flashcards, and quizzes.
• Engineered fault-tolerant ingestion handling 5 distinct transcript failure modes with per-video isolation.
• Multi-model routing layer across OpenRouter and Gemini with a 3-stage hallucination filtering pipeline.
• Stack: Python, FastAPI, React 19, Vite, Tailwind CSS, Azure Container Apps, Docker, In-Memory Vector Search.

LLM-to-IR Architecture · Autonise Internship (May–June 2026)

• Decoupled semantic reasoning from coordinate emission by creating a 16-tool domain-specific language (calculate_intersection, tangent_to_circle, compute_centroid, etc.).
• Enforced a strict tool-before-coordinate mandate so the model outputs a topological constraint graph rather than pixel coordinates.
• Deterministic NumPy + Matplotlib compiler analytically solves the equations into clean vector SVGs, dropping diagram failure rates from ~60% to 0%.

High-Density Document Segmentation & OCR

• Fine-tuned YOLOv10m on a 33k-image synthetic dataset with variable fonts, multi-column layouts, and margin shifts (96.1% mAP50 / 88.9% mAP50-95).
• Diagnosed and fixed an upstream float16 export defect that silently zeroed out Distribution Focal Loss (DFL) gradients.
• Integrated into a two-collection MongoDB staging pipeline with scope-enforced reads and interactive human review.

Multi-Agent State Machines & Tool Coordination

• Implemented LangGraph state machines using TypedDict state, conditional router dispatching, and looping control flow.
• Built 8 Google ADK agent patterns including Sequential, Parallel, Session-persistent, and Tool-calling agents.
• Stack: Python, LangGraph, Google ADK, Streamlit.

Attention-Based Multimodal Image Captioning

• ResNet-50 feature extractor paired with an LSTM language decoder augmented with Bahdanau additive attention.
• Generates dynamic alignment heatmaps over image regions during autoregressive caption generation.
• Stack: PyTorch, Torchvision, Python, Bahdanau Attention, ResNet-50.

• AI/ML Team Lead @ IIT Bombay: Leading an engineering cohort preparing for and competing in national and global AI/ML competitions, data challenges, and hackathons.