About Me
I'm an Applied AI Engineer based in India, focused on building LLM systems that work in production — not just in notebooks.
I've deployed offline GenAI inside DRDO's air-gapped defense network, built voice AI agents for enterprise clients, and implemented LLaMA 3 from scratch to understand transformers at the architecture level.
I care about the gap between research and deployment — and closing it.
What I can build for you
- RAG & document intelligence systems — I build pipelines that let your product answer questions over your own data. PDFs, DOCX, databases — ingested, embedded, and queryable. Shipped for production, not just demos.
- Voice AI agents — end-to-end conversational agents for inbound and outbound calls, with configurable tone, persona, and call flow logic. Already deployed for 3+ enterprise clients via ElevenLabs + custom orchestration.
- LLM integration & automation — connect LLMs to your existing workflows. Prompt engineering, structured output, tool use, evaluation frameworks. I've done this in air-gapped defense environments — your stack is easier.
- Full AI product builds — if you have an idea and need someone to own the AI layer end-to-end: architecture, backend, and deployment. I work best with early-stage startups who move fast.
- Proven, not theoretical — every service above has a shipped project behind it. DRDO, Leapon, Giggr.
I don't pitch technology I haven't deployed.
Skills
- Large Language Models (LLMs)Deployed in production at DRDO and Leapon
- RAG ArchitecturesShipped multiple production systems
- Generative AIUsed daily across all recent roles
- Prompt EngineeringCore part of every LLM project
- Fine-Tuning (LoRA / PEFT)Applied on HuggingFace models
- Transformer ArchitectureImplemented LLaMA 3 from scratch
- NLPUsed across multiple projects
- LangChainPrimary orchestration framework
- OllamaUsed for local LLM deployment at DRDO
- HuggingFace TransformersModel loading, fine-tuning, inference
- PyTorchBuilt models and architectures from scratch
- TensorFlowUsed in early CV and research projects
- FastAPIBackend for AI product APIs
- DjangoProduction backend at Leapon
- FAISSUsed in GenieFile RAG pipeline
- Neo4jDeployed Knowledge graph in Giggr and other projects
- MySQLProduction DB at Leapon
- PineconeUsed in LLM QA Chatbot project
- ElasticsearchBuilt 1.3M+ institute API at Giggr
- Google Cloud PlatformMigrated full production stack to GCP
- Amazon Web ServicesUsed across multiple production systems
- GitHub Actions (CI/CD)Used at Leapon and Giggr
- Microsoft AzureUsed at Maersk for ML pipelines
- PythonPrimary language across all roles
- JavaScriptUsed at Giggr and portfolio
- SQLProduction queries at Leapon and Maersk
- JavaAcademic and systems work at Maersk
- C++Academic background
- Pandas / NumPyUsed in every data project
- Deep LearningCNNs, autoencoders, Siamese nets
- Computer Vision (OpenCV, YOLO)87.4% accuracy on YOLO at Giggr
- scikit-learnClustering and ML models at Maersk
- PySparkLarge-scale data at Maersk
- MLflowExperiment tracking
Experience
Defence Research and Development Organisation (DRDO)
January 2025 – August 2025
Graduate Apprentice (CSE)
- Air-Gapped LLM Deployment: Deployed LLaMA and Mistral inference stacks across Linux and Windows in a high-security network with no internet access — enabling 12+ researchers to run AI queries locally.
- Research Efficiency: Reduced average query turnaround from ~2 hrs to under 35 min by designing a localized model execution pipeline that eliminated all external dependencies.
- Mentorship: Guided and mentored a group of 10 interns through their internship project — overseeing technical direction, troubleshooting blockers, and ensuring delivery within the program timeline.
- Automated Tender Generation: Built a Tesseract + EasyOCR extraction pipeline feeding a LangChain/Ollama RAG system that outputs structured LaTeX tender documents from noisy scanned PDFs — eliminating manual drafting for 20–30 documents/week.
Leapon
February 2024 – Present
Product Developer (concurrent freelance)
- Voice AI Agent: Built a prompt-orchestrated calling agent using ElevenLabs, enabling 3+ enterprise clients to automate inbound/outbound call flows with configurable tone, persona, and conversation logic.
- Cloud Migration: Migrated full production stack from AWS (RDS MySQL + S3) to GCP with under 60 min of downtime — cutting monthly infrastructure costs by 46% ($130 → $70/month).
- OCR Automation: Automated business card digitization using Google Vision API, parsing card images into structured vCard and Excel output for a sales team's CRM onboarding process.
Giggr Technologies
June 2023 – February 2024
AI/ML Engineer
- Event AI System: Built a system integrating facial recognition, object detection, and GPS tracking — achieving 87.4% accuracy using OpenCV, YOLO, and DeepFace.
- Public API: Delivered an Elasticsearch-powered API covering 1.3M+ Indian education institutes via web scraping and AWS Shell Scripting.
- Conversational AI: Built a Dialogflow + GPT chatbot with Firebase integration, growing user engagement from ~800 to ~1,500 sessions/month within the first quarter.
Leapon
February 2023 – June 2023
Backend Developer
- Django REST API: Built a customized backend for a platform with 50+ advisors, with unit testing, bug fixing, and CI/CD integration. Stack: Python/Django, SQL, AWS, Git, JIRA.
Maersk Global Service Centres
July 2022 – February 2023
Data Science Intern
- Port Clustering: Clustered global port data using Gaussian Mixture Models for cost estimation — stack: PySpark, Databricks, Microsoft Azure.
- Logistics Forecasting: Modeled container turn times and forecasted attachment ratios using XGBoost and Regression, contributing to a 15% increase in freight profit.
- Customer Segmentation: Applied K-Means clustering to segment large-scale customer datasets, identifying behavioral patterns to optimize digital acquisition strategies.
Projects
Featured Projects
RAG document assistant combining FAISS vector search with a Neo4j Knowledge Graph for cross-document entity reasoning. Includes a custom eval framework tracking hallucination rate and retrieval precision.
Full LLaMA 3 architecture built in PyTorch from scratch — tokenization, RoPE, RMSNorm, multi-head attention, feedforward layers — verified against Meta's official weights. No HuggingFace abstractions.
Siamese Neural Network with ResNet backbone and spatial attention to distinguish genuine vs. forged signatures. End-to-end pipeline with contrastive loss training — built for fraud detection and identity verification.
Other Projects
LoRA + PEFT fine-tuning on HuggingFace models for optimized text generation.
Streamlit RAG chatbot for real-time IPO financial analysis using Pinecone vector DB.
BERT + Gemini-Pro pipeline for ATS compatibility scoring and resume recommendations.
Multi-model CV system for gastrointestinal disease detection using Detectron2 and MaskRCNN.
Django + ResNet50 web app for real-time object identification from uploaded images.
Blog content generator via speech recognition and Google GenAI with tone customization.
AI-driven document summarization tool for bloggers built with Gemini and Streamlit.
CNN, ANN, RBF architectures implemented from scratch to understand mathematical intuition.
ResNet18 fine-tuned on AWS SageMaker to classify 133 dog breeds with hyperparameter profiling.
Transfer learning model deployed on SageMaker achieving 83.16% furniture classification accuracy.
CNN classifier for cats vs dogs using Keras Sequential API achieving 98.7% accuracy.
Speech recognition assistant for music playback, Wikipedia lookups, and system tasks.
Publication
Research framework for HSI classification using a CNN bottleneck AutoEncoder with Hyper Spectral Net. Achieved 0.998 accuracy on classification. Published in IJCISIM.