Designing and building AI systems, from model inference and optimization to production-facing applications.
Bengaluru, India · Open to AI / ML engineering opportunities
I build AI-powered applications and developer tools, with particular interest in LLM inference, memory-efficient model execution, computer vision, and reliable backend services.
My work spans machine learning experiments, application architecture, model integration, API development, and deployment. I'm particularly interested in making advanced AI systems more practical on limited compute.
FOCUS LLM inference · Applied ML · Computer vision
ENGINEERING Python · PyTorch · TensorFlow · FastAPI
BUILDING AI developer tools · Intelligent applications
INTERESTS Memory optimization · Agent systems · Deployment
Explore repository · PyPI package
A Python package designed to run large language models with constrained GPU memory by streaming transformer layers between host memory/storage and GPU VRAM rather than keeping the entire model resident in VRAM.
- Layer-wise execution with asynchronous prefetching and double buffering
- Dedicated CUDA compute and transfer streams with event synchronization
- Pinned host memory, reusable GPU scratchpad buffers, and KV-cache budgeting
- Python SDK and CLI for model installation, interactive inference, and benchmarking
Stack: Python PyTorch CUDA LLM Inference GPU Memory Optimization
Designed for resource-constrained inference. See the repository for supported models, requirements, benchmarks, and current limitations.
flowchart LR
A["Model weights<br/>RAM / disk"] --> B["Pinned host memory"]
B --> C["CUDA transfer stream"]
C --> D["GPU scratchpad<br/>Double buffer"]
D --> E["CUDA compute stream"]
E --> F["Transformer output"]
G["KV-cache budget"] -.-> D
C -. "Prefetch next layer" .-> D
Measured on a 4 GB VRAM GPU. The reported timings compare sequential layer transfer/execution against asynchronous prefetching in a streaming microbenchmark; they are not full-model tokens/second measurements.
| Layers | Sequential | Prefetch | Speedup | GPU scratchpad |
|---|---|---|---|---|
| 8 | 68.08 ms | 56.70 ms | 1.20× | 128 MB |
| 16 | 136.51 ms | 126.11 ms | 1.08× | 128 MB |
| 24 | 219.05 ms | 170.54 ms | 1.28× | 128 MB |
| 32 | 278.85 ms | 265.03 ms | 1.05× | 128 MB |
Benchmark details and methodology · Quickstart
A deepfake detection application for analyzing image and video content, organized as a FastAPI backend and React/TypeScript frontend, with Docker-based local setup.
Stack: Python FastAPI React TypeScript Computer Vision
A code review tool combining deterministic Python AST/static analysis with LLM-assisted reasoning to identify issues and produce actionable review feedback.
Stack: Python FastAPI Groq API LLMs React
A computer vision system with real-time face recognition, MobileNetV2-based learning, liveness detection, and attendance tracking.
Stack: Python TensorFlow OpenCV MobileNetV2
An intelligent inbox monitoring project that combines email integrations, importance scoring, and WhatsApp notifications for priority messages.
Stack: Python Gemini API Integrations Automation
| Domain | Tools and technologies |
|---|---|
| AI & deep learning | Python, PyTorch, TensorFlow, scikit-learn, OpenCV |
| LLM engineering | Transformer inference, CUDA streaming, model optimization, LLM APIs |
| Backend engineering | FastAPI, Flask, REST APIs |
| Applications | React, TypeScript, JavaScript, HTML, CSS |
| Data & development | NumPy, Pandas, Git, GitHub, Docker |
Crismatec — Software Development / AI Integration
December 2024 – Present
- Contributed to a Student Learning Portal, developing interfaces and connecting them to backend APIs.
- Worked on integrating AI-powered learning workflows, including personalized learning support, quiz generation, and assessment-related features.
- Collaborated with the development team on integration testing, troubleshooting, and application improvements.
Interested in the intersection of applied AI, software engineering, and resource-efficient inference systems.
Inference performance — reducing peak GPU memory requirements and studying latency / throughput tradeoffs.
Practical AI systems — integrating models into useful, maintainable applications with APIs and clear interfaces.
Agentic software — exploring tool-using AI workflows, automation, and developer productivity.
For implementations, experiments, and development history, visit the repositories tab. Selected work is highlighted above.