Headline

AI Engineer @ Centific | Building & Fine-tuning LLMs, RAG, Agentic AI Systems | Deep Learning Researcher | GenAI · Multimodal · Computer Vision · Model Optimization

About

AI Engineer & Deep Learning Researcher. 5+ years turning research into production AI systems. I work at the intersection of applied research and deployed systems — publishing in peer-reviewed venues while shipping models that run in production at scale. AT CENTIFIC, I design and deploy AI systems across LLMs, computer vision, and multimodal learning, including: - Redesigned a law enforcement evidence-review pipeline (video, audio, documents), cutting officer review time ~85% by combining deterministic techniques with narrowly-scoped reasoning-model calls, deployed on-premise for data privacy. - Built Spanglish (Spanish-English code-switched) ASR for law enforcement, reaching 0.1 WER through synthetic data generation and domain-specific fine-tuning. - Designed SurgQ-VLM, a two-stage SFT → GRPO surgical video QA model with a novel dual-stream compressor, evaluated on a sealed 21,840-question benchmark. - Built a real-time multi-camera security analytics pipeline combining rule-based alerting with VLM judgment, reaching 89.5% correct aggression/weapon alerts in live evaluation. Before Centific, at Ignitarium and Qualitas Technologies, I worked on 6DoF pose estimation for robotics, edge-device classification models, fault diagnosis from sensor data, and defect detection — optimizing for cloud, edge, and embedded deployment with TensorRT, TFLite, and mixed precision. Alongside industry work, I've been an active researcher through IIT Roorkee's CANDLE Lab, with peer-reviewed publications in venues including WACV, ICDAR, ISQED, and Elsevier, spanning low-light image enhancement, scene text understanding, PCB defect classification, and nuclei instance segmentation. I'm looking to go deeper into research-driven roles — Applied Scientist / Research Scientist positions — where I can push on LLM alignment (RLHF/GRPO), agentic systems, and multimodal reasoning, while staying close to real deployment constraints. Open to connecting on AI research, applied deployment, or collaboration — reach out here or at dmakwana503@gmail.com.

Location

Hyderabad

AI stack

anthropic apis

claude code

claude

hugging face

langgraph

ChatGPT Images

Work experience

AI Engineer at Centific, Mar 2024 - Present · 2 yrs 7 mos, Hyderabad, - Redesigned a law enforcement case-review pipeline for video, audio, and document evidence, cutting officer review time ~85% (75-90 min → 12-15 min) by combining deterministic techniques (timestamp OCR, identity tracking, native-text parsing) with narrowly-scoped reasoning-model calls, deployed on-premise for data privacy compliance. - Built Spanglish (Spanish-English code-switched) speech recognition for law enforcement, reaching 0.1 WER by generating synthetic conversational training data from expert-identified critical word lists and pretraining on novel concatenated English-Spanish corpora before domain-specific fine-tuning. - Built a real-time multi-camera security video analytics pipeline combining deterministic rule-based alerting with scoped VLM judgment and a correlating decision-layer agent, achieving 66.7% recall at 80% precision on unattended-package detection and 89.5% correct aggression/weapon alerts in live evaluation. - Designed and trained SurgQ-VLM, a two-stage (SFT → GRPO) surgical video QA model with a novel bidirectional question-conditioned dual-stream compressor, reaching 0.779 keyword accuracy and a 10x BLEU-4 gain over zero-shot on a sealed 21,840-question benchmark. - Built an LLM-powered data structuring pipeline for a food delivery client using GPT-4 and LangChain to normalize raw product data into schema-compliant structured outputs, with a GPT-4-as-judge quality gate that auto-approved ~90% of rows and reduced manual review effort by 85%.

AI Engineer at Ignitarium, Dec 2021 - Jan 2024 · 2 yrs 2 mos, Bengaluru, - Developed a 6DoF pose estimation pipeline for robotic arm pickup of small industrial parts, resolving severe data scarcity via 3D mesh rendering for synthetic training data generation, multi-class YOLOv8 detection, and MegaPose for pose estimation. - Developed and implemented high-performance computation kernels, reducing system latency in data transformations and computer vision tasks, improving processing speeds and system efficiency by 2x. - Built a lightweight multitask CNN for gender (binary) and age-group (4-class) classification on a custom edge device, combining multiple Asian face datasets to overcome data scarcity and achieving 94% accuracy. - Designed a prototype solution for fault diagnosis from machinery sensor signal data, achieving 98% accuracy for fault classification utilizing Fourier and wavelet-based noise reduction with LSTM and dilated convolution layers.

Research Intern at IIT Roorkee — CANDLE Research Lab (with Dr. Sparsh Mittal), Nov 2021 - Oct 2024 · 3 yrs, Roorkee, Conducted research across computer vision and multimodal learning, contributing as first, second, or third author across 5+ projects resulting in publications at WACV, ICDAR, ISQED, and Elsevier venues. - Led research on PCB component detection/segmentation for recycling and low-light image enhancement as first author — defined the research problem, reviewed prior work, designed and implemented custom modules to address their limitations, and ran experiments on the FICS PCB dataset and LOLv1 dataset. - Contributed as second author to research on scene-text removal and PCB component classification. - Contributed to global Fourier self-attention for multimodal downstream tasks and nuclei instance segmentation as third author, including experiments on prior methods and data/post-processing pipelines. - Co-authored paper writing and prepared responses to anticipated reviewer questions across multiple submissions.

Associate Machine Learning Engineer at Qualitas Technologies Pvt Ltd, Jan 2021 - Dec 2021 · 1 yr, Bengaluru, - Developed and implemented Faster R-CNN and SSD-based object detection models, achieving more than 95% mAP for chassis number detection. - Implemented DeepLabV3 segmentation for manufacturing defect detection and optimized inference using TensorRT, OpenVINO, mixed precision, and INT8 quantization, achieving ~35% latency reduction with less than 1% accuracy drop. - Reduced manual image annotation effort by ~90% by designing and building an active learning pipeline (Faster R-CNN pseudo-labeling + confidence filtering) for manufacturing inspection data, cutting labeling time from ~83 hours to ~8 hours per 10K-image project.

Machine Learning Intern at Qualitas Technologies Pvt Ltd, Sep 2020 - Jan 2021 · 5 mos, Bengaluru, - Designed a novel and custom computer vision architecture and developed data‑memory efficient pipelines for computer vision/deep learning‑based applications, utilizing advanced techniques and industry best practices to optimize system performance and enhance efficiency.

Education

Bachelor of Engineering - BE, Computer Science, LJ Institute of Engineering and Technology

Public links

github: https://github.com/dhruvmakwana

dhruvmakwana.github.io: https://dhruvmakwana.github.io/