AI systems with research depth and product intent.

AI Engineer, Vizit

Boston, MA

• Built Cosmo, a zero-shot product taxonomy classifier mapping retailer titles to 3,700 Amazon taxonomy nodes using CLIP text embeddings, z-score-calibrated cosine similarity, and Walmart-aware calibration; automated client onboarding and unlocked Vizit's expansion from Amazon-only to Walmart.
• Productionized Cosmo on AWS Lambda with FP16 ONNX compression from 1.71 GB to 248 MB, SHA256-pinned artifact bundles, atomic five-minute rollback, nightly canary stability checks, and Datadog monitoring for confidence, drift, no-match rate, and latency.
• Built Cleo, a 14-class product image classifier fine-tuned from CLIP ViT-L/14 on 912K labeled examples through Databricks pipelines; reached 87.2% accuracy and 84.1% macro F1, beating the prior production model by 3.15 accuracy points.
• Developed ACI, a calibrated regression system linking image features to Amazon conversion and search-impact signals across 11,269 ASINs and 300+ benchmarks, redirecting roughly $400K per year in creative spend toward higher-ROI work.
• Designed and implemented Project Spark, a multi-agent recommendation and diagnostic system combining predictive scoring, benchmark retrieval, interpretability modules, and GPT-4.1 plus Claude reasoning; improved visual appeal scores by 30% across 20+ global brands.
• Built image interpretability tooling with PCA-based eigen-image decomposition, Grad-CAM, and FastAPI delivery; also deployed creative-intent classifiers, hierarchical taxonomy systems, and a 5M+ asset semantic search engine with drift-aware monitoring.

AI Engineer Intern, AG Mednet Inc.

Boston, MA

• Built Judi-Insights, a clinical text-to-SQL assistant over trial databases using LangChain, Llama-3-70B on AWS Bedrock, and a React frontend; reduced report generation time from four hours to 25 minutes for 50+ clinical researchers.
• Benchmarked Llama-3-70B against Claude 3.5 Sonnet for complex multi-table SQL generation, reaching 80% accuracy on the clinical analytics workflow.
• Built a clinical RAG pipeline with FAISS, parent-child chunking, and cross-encoder reranking, reducing hallucination risk from 18% to under 5%.
• Optimized search and retrieval across 10M+ clinical records using composite indexes, materialized views, query-plan analysis, and dynamic view triggers, reducing p95 latency from 800 ms to 180 ms and supporting 500+ daily researcher queries.
• Deployed production ML workflows on AWS EC2 with Docker, GitHub Actions CI/CD, TestRails-based evaluation tracking, and FDA-aligned compliance processes.

Head of Artificial Intelligence Division, Expand-AI

Bangalore, KA

• Designed the complete data annotation pipeline for computer vision and NLP tasks.
• Developed large ViT-based segmentation systems and paired text plus visual encoders for class-definition-to-annotation workflows.
• Benchmarked against DETR, YOLO, OneFormer, Swin Transformers, and related architectures.
• Built AWS EC2 and Fargate production pipelines, visualization dashboards, serving scripts, retraining flows, and API integrations.

Deep Learning Researcher and Incubatee, IIT Hyderabad

Hyderabad, TS

• Focused on practical deep learning applications in healthcare, navigation, and media.
• Worked with stakeholder pain points to design and test end-to-end adaptive AI systems.

NLP Researcher, The Nirvana Project, RAPYD.AI

Hannover, Germany

• Built benchmarking systems for the Nirvana summarization architecture against T5, DeBERTa, RoBERTa, and related baselines.
• Curated large podcast audio and text datasets and built the full podcast-to-summary application pipeline.
• Applied pretraining, speaker recognition, conditional generation, and quantization to improve quality and efficiency.

Graduate Research Assistant, IoT Innovation Lab

Northeastern University, Boston, MA

• Led a Digital Twin initiative that personalizes LLM-based avatars using behavioral signals extracted from LinkedIn, Instagram, and related platforms.
• Built multi-source pipelines, LoRA adapters, and unified personality embeddings for context-aware language generation.

Graduate Research Assistant, Northeastern University

Boston, MA

• Researched multimodal attention architectures for modeling commercial success across 150K videos using video, audio, text, and attention signals.
• Developed context-based scene-to-text and audio-to-scene retrieval approaches for richer video understanding.

Research Head, AIR Center, VIT-AP University

Amaravati, AP

• Led cybernetics-inspired R&D projects involving human feedback loops for adaptive machine intelligence.
• Worked in collaboration with researchers connected to the University of North Texas and the University of South Florida.

Deep Learning Researcher, Centre for Research and Innovation in AI, IIT Hyderabad

Hyderabad, TS

• Developed EMG technology to track and isolate hyperlocal muscle impulses during involuntary movement.
• Designed LiDAR-based scene mapping and pheromone-inspired AutoNAV systems for TiHAN testbeds.
• Researched latent space manipulation for Depth2Mesh and diffusion-based text-to-scene virtual production and lip syncing.

Cosmo

Zero-shot product taxonomy classifier using CLIP text embeddings, centroid scoring, and z-score calibration across 3,700 taxonomy nodes; automated onboarding and enabled Walmart expansion.

Cleo

14-class production image classifier fine-tuned from CLIP ViT-L/14 on 912K labeled examples, with controlled experiments across LoRA, contrastive, and late-fusion variants before selecting the winning full fine-tune.

ACI

Calibrated regression system connecting image features to conversion-rate and search-impact signals, chosen over LightGBM and small MLP alternatives for stability and interpretability.

Captioning Pipeline

Generated 69,497 captions across 14 product-image classes using Gemini Flash and Flash-Lite with prompt iterations, anti-hallucination constraints, and analysis tooling; contrastive fine-tuning experiments regressed, so the cleaner caption-derived prototype approach won.

DiffusionSync

Diffusion models for realistic lip syncing from arbitrary face angles through latent space manipulation rather than GAN-only mouth patch generation.

Retractive Path Navigation

Pheromone-inspired Q-learning and federated learning framework for efficient navigation with lower compute cost and higher routing efficiency.

Gargantua

Real-time EEG classification and recognition using a reconstructed VGG-style neural network with visual heatmap outputs.

CLEO

Smart glasses for alcohol and cigar monitoring using custom YOLO models, vitals tracking, and medical alerting.

MaLang

Decentralized deep intelligence system to tackle toxic text across devices using contextual and demographic-aware detection layers.

Dexterat

Neural style transfer on video using VGG-19 and optical flow for smooth frame interpolation and feature preservation.

Face Editing for Indian Diaspora

Fine-tuned latent manipulation workflows on Indian faces for more consistent and culturally relevant face editing outputs.

VISU

3D printed humanoid robot supporting face recognition, voice interaction, posture tracking, movement replication, and surveillance.