Bridging Computer Vision & LLMs in Production
I've trained CNNs to see and transformers to reason. Now I build systems where both work together.
Figure 1. Flow matching trains a network to give every sample a velocity that carries it from noise to the data in a straight line, xt = (1 − t)·x0 + t·x1. Generating is just following the arrows from t = 0 to 1. Here the data is a stippled portrait of Gradi.
01Abstract
I'm an electrical engineer who ended up in AI. After clearing GATE and BARC, I chose research over a government job — which took me to IISc Bangalore and my first paper, on exemplar‑free continual learning.
Since then I've worked where models meet users: diffusion pipelines that shipped photorealistic product imagery at Avataar AI, and now agentic AI systems at Armada AI, where I own retrieval, evals and observability end to end.
The boundary between what machines see and what they understand keeps blurring. I build at that edge, and I still publish — most recently on fine‑grained attribute control in diffusion‑based visual storytelling.
02Work Experience
AI Engineer
Armada AI
- Designed the shared CRAG + controllable ReAct architecture behind every knowledge‑base agent, cutting pipeline cost to ~1/3 of baseline by deduplicating reformulation and embedding calls into one graph node
- Built a two‑tier eval framework: retrieval‑confidence and CRAG‑score gating before generation, RAGAS groundedness scoring after, against hand‑curated ground truth
- Built an online‑learning BM25 index with incremental update and unlearn, closing organizational‑vocabulary gaps that dense retrieval missed
- Replaced single‑pass image captioning with a two‑stage classify‑then‑caption pipeline, fixing the flowcharts and screenshots it failed on
- Replaced per‑format ingestion pipelines with a normalized‑markdown layer, cutting new‑format onboarding effort by ~50%
Research Engineer
Avataar AI
- Built the lifestyle image generation pipeline on Flux and ControlNets
- Modified diffusion sampling with intrinsic decomposition for better object reconstruction
- Built classification for low‑data scenarios with CLIP, BLIP2, and Qwen2.5
- Improved segmentation accuracy with BiRefNet and SAM + YOLO‑world
Teaching Assistant
IISc Bangalore - Signal Processing, Digital Image Processing
- Built reference implementations for continual learning (L2P, DualPrompt) and self‑supervised pretraining (MoCo, SimCLR)
- Designed and graded assignments across both courses, from frequency‑domain filtering and SIFT to deep classification
03Research
Academic Service
04Projects
Agentic RAG System
Armada AI
Multi‑agent RAG pipeline with LangGraph orchestrating planner, retriever, evaluator & generator stages. Hybrid search via dense + BM25 + RRF over Qdrant with Jina reranking.
LangGraphQdrantHybrid SearchFastAPIDocker
Lifestyle Image Generation
Avataar AI
Product image generation pipeline on Flux + ControlNets. Modified diffusion sampling with intrinsic decomposition for photorealistic object reconstruction.
FluxControlNetDiffusionSAMCLIP
05Skills
AI/ML
LLMsRAGCRAGAgentic AIComputer VisionDiffusion ModelsNLPHybrid SearchBM25Evals (RAGAS)
Frameworks & Tools
PythonPyTorchLangGraphLangChainHuggingFaceFastAPIChainlitPlaywrightDocling
Infrastructure
DockerHelmAWSPostgreSQLQdrantLangfuseSSO/RBAC
06Education
M.Tech in Artificial Intelligence
Indian Institute of Science (IISc), Bangalore
2022 - 2024 · CGPA: 8.0/10.0
Pattern Recognition & Neural Networks, Computer Vision, Digital Image Processing, Deep Learning for NLP, LLMs for Practical NLP, Stochastic Models, Optimization
B.Tech in Electrical Engineering
Bhagalpur College of Engineering, Bhagalpur
2018 - 2021 · CGPA: 8.75/10.0
Diploma in Electrical Engineering
Government Polytechnic Muzaffarpur, Muzaffarpur
2015 - 2018 · 77.73%
Achievements & Certifications
Rank 1
BCECE LE 2018
State Lateral Entry Exam
07Contact
Let's Build Something Together
Looking for collaboration on AI/ML projects, research opportunities, or just want to chat about generative models and agentic systems.