DeGirum

AI Engineer

Manning Publications Co.

Technical Author

traversaal.ai

- Building End to End enterprise level LLM powered Co-pilots and RAG systems.
- Deploying Opensource LLMs such as Mistraal on Multi-GPU systems 
- Utilizing Huggingface frameworks such as Deepspeed, DPO, RLHF and Bits&Bytes for finetuning LLMs

AI Researcher Co-op

Packt

Technical Reviewer

Santa Clara University

Role and Project:

Position: Research Assistant at Santa Clara University.
Project Goal: Automate fetching insights from syllabus PDF to help faculty in optimizing syllabus planning to maximize student performance through better syllabus structuring.

NLP Techniques:
Implemented NLP techniques.
Tasks: Entity recognition, keyword extraction, text embedding.

LLMs Integration:
- Finetuned Llama-2 and Falcon 7B LLMs using PEFT and QLora on custom dataset
Purpose: Enhance syllabus insights and correlate with student performance.

Research Assistant - TailorED Project

Teksun Inc

1. Role and Responsibilities:
   - Position: Machine Learning Engineer at Teksun Microsys Pvt Ltd.
   - Focus: Developing and implementing end-to-end AI and ML pipelines.

2. Lane Detection Algorithm:
   - Developed a State of the Art lane detection algorithm for an autonomous assisted driving system.
   - Techniques Used: Gradient filter, perspective transform, and custom image segmentation in PyTorch.
   - Result: Achieved a 96% pixel accuracy for assisted driving mode on Indian roads.
   - Client: Japanese Car manufacturing company with 80,000+ customers in India.

3. Video Classification Model:
   - Designed a lightweight video classification model.
   - Architecture: CNN-LSTM.
   - Application: Accurate detection and counting of repetitions for 40 different gym exercises.
   - Deployment: Implemented and deployed on Nvidia Jetson nano.

4. Object Detection for Hazard Detection:
   - Implemented YOLOv5 Object Detection for hazard detection in commercial buildings.
   - Contribution: Played a role in building incremental learning pipelines.

5. ML Orchestration and Monitoring:
   - Implemented ML orchestration and monitoring using Airflow DAGs.
   - Application: Qualcomm 625 based commercial product.
   - Clientele: Served numerous B2B clients in India and the US.

Machine Learning Engineer L2

HypeTeq Software Solutions Pvt. Ltd

Role and Responsibilities:

Position: Junior Software Engineer (AI/ML) at Hypeteq Software Solutions Pvt. Ltd.
Contribution: Actively involved in the development of energy prediction systems.
Wind Energy Prediction Project:

Project Focus: Wind energy prediction.
- Implemented custom Time-series based stacked machine learning boosting model.
Performance: Achieved a remarkable Mean Absolute Percentage Error (MAPE) of 1.40%.
Benefit: Improved resource utilization in the energy production process for seimens based wind-farms across Western India.

Junior Software Engineer - ML/AI

Software Engineer trainee

Udemy

I am an online ethical hacking trainer at Udemy learning platform and share my hacking skills that are highly in demand. I am on my way to make technology a better and a safer thing to dwell with.

Online Instructor

Authored a book on ethical hacking and cyber secuirty that aims to make technology a better and a safer thing to dwell with by spreading awareness about cyber crimes and threats to information assets.

Book Author - "Hack yourself before someone else does"

As a passionate AI practitioner, I have a strong hold over various AI frameworks, including NLP, computer vision, and machine learning. With over five years of industry experience as an AI Engineer in both India and the United States, including roles at early-stage startups in Silicon Valley, I excel at bridging business challenges with data science methodologies. My expertise lies in connecting business problem with comprehensive end-to-end AI solution architectures and executing them effeciently.

I am a tech author and speaker. I have been invited to speak at World's No. 1 MSBA course at UCLA. I was also invited to judge AI Hackathon at UC Berkeley. I spoke at events like MLOps Community event and Founders tech weekend about LLMs and Enterprise search. I am currently authoring a book on Enterprise RAG with Manning publications. I have also served as a teaching assistant on Maven's top selling LLM course with Prof. Hamza Farooq.

Key areas of experience:

- LLMs : Huggingface Transformers, Mistral, Llama, OpenAI, RAG, VectorDB, Langchain, LlamaIndex

- Computer vision : Tensorflow, PyTorch, YOLO, pose estimation, object detection, CNN-LSTM, GANS, Pix2Pix models

- Edge AI : OpenVino toolkit, Nvidea ONNX, deploying models on Rpi and Jetson Nano

- NLP : LLMs, Abstrsctive Text summarization, Entity extraction - NER, SPACY, BERT, Hugging face models

- Timeseries models : ML based stacked regression model, Adaboost, XGBoost, SVR

- Cloud : Deploying models on AWS EC2, building ECS clusters, load balancing using AWS ELB 

- Model deployment : Flask web apps, fastAPI, Streamlit, edge/cloud, Gradio.

Since last few weekends, I have been working on building something that powers most successful products across domains.

Introducing RecommendKit: Production-Grade Recommendation Systems Made Accessible

Github: https://lnkd.in/dw38Yu84

Recommendation systems are backbone of some of the most successful products today—from Instagram's Explore page to Amazon's product suggestions, from Netflix's content discovery to LinkedIn's job matching. Yet despite their widespread use, there's a surprising gap: no ready-to-use, out-of-the-box package that handles multi-modal data (images, text, temporal sequences, categorical features) with true production-grade scalability.
Most existing libraries either focus on single modalities, require extensive custom code for real-world data, or can't handle the cold-start problems that plague production deployments. I wanted to change that.

RecommendKit is a comprehensive Python package designed to democratize access to Instagram-scale recommendation technology. Here's what makes it different:

- True Multi-Modal Support: Seamlessly combines 5 data types—images, text, categorical, continuous, and temporal features—in a single unified architecture

- Hash-Based Categorical Encoding: Handles unseen categories without model retraining—a critical feature for production that solves the cold-start problem other libraries ignore

- Flexible Architecture: Switch between SimpleFusion (fast, stable) and Attention-based fusion (complex interactions) with a single config change—no code modifications needed

- Intuitive JSON Format: Simple data structure that anyone can understand. No complex preprocessing pipelines or data engineering overhead

- Production-Ready: Complete training and inference pipelines, batch embedding generation, vector database integration (Qdrant), and ANN search for real-time recommendations

- Truly Extensible: Modular design with plug-and-play encoders. Add custom components without touching core code.

The system automatically handles positive/negative sampling, supports distributed training, GPU acceleration, and is designed to scale from thousands to millions of interactions.

This is Just the Beginning. Phase 1 is complete, but there's much more to build:

- Advanced ranking algorithms and loss functions
- Multi-task learning capabilities
- Real-world multi-modal dataset testing and benchmarking
- Performance optimization for billion-scale deployments
- Domain-specific encoder packages
- Production deployment tools and monitoring

Looking for Collaborators:
If you're passionate about ML infrastructure, recommendation systems, or democratizing AI technology, I'd love to have you join this journey.

Let's build something that makes production-grade recommendation systems accessible to everyone. Drop a comment or DM if you'd like to collaborate!

Darshil Modi

Qdrant

Creator of RecommendKit, AutoMeta RAG framework | Multimodal AI Engineer | AI Agents | Tech speaker | Published Author - “Enterprise RAG” @ Manning | Open source contributor

Boost Hailo Inference Performance with Eager Batching

Curious how batching impacts throughput on Hailo-8 and Hailo-8L using DeGirum #PySDK? This guide walks through benchmarks, model context behavior, and when eager batching makes a real difference.

👉 Learn how to get the most out of your Hailo accelerators with DeGirum Tools and PySDK: https://lnkd.in/dAE4qBng

taplio

Darshil Modi's Linkedin Analytics

Darshil Modi

Performance ScoreDarshil Modi's LinkedIn Performance

Darshil Modi's Best Posts (last 30 days)

Ready to grow your LinkedIn presence?

taplio

.css-6ytdur{background-image:linear-gradient(284deg, rgba(39, 43, 72, 1) 0%, rgba(33, 47, 183, 1) 100%);color:transparent;-webkit-background-clip:text;background-clip:text;}Darshil Modi's Linkedin Analytics

Darshil Modi

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Darshil Modi's LinkedIn Performance

Darshil Modi's Best Posts .css-1emoo84{font-size:var(--chakra-fontSizes-md);font-weight:400;color:var(--chakra-colors-gray-600);}(last 30 days)

Ready to grow your LinkedIn presence?

Darshil Modi's Linkedin Analytics

Performance ScoreDarshil Modi's LinkedIn Performance

Darshil Modi's Best Posts (last 30 days)