Hello, I'm

Tuhin Acharjee

I build |

ML & Computer Vision Engineer with 3.5+ years of experience building real-time video analytics, edge AI, and production deep learning systems.

0 Years Experience
0 Detection Accuracy
0 Real-time Processing
0 Efficiency Gain
View My Work Download Resume
Scroll

About

Building intelligent systems
that see and understand

Machine Learning and Computer Vision Engineer with 3.5+ years of experience building real-time video analytics, edge AI, robotic vision, face recognition, and Python backend systems.

I specialize in taking ML models from research to production — optimizing inference pipelines, deploying on edge devices with NVIDIA DeepStream and TensorRT, and building the backend services that make AI systems actually useful.

My work spans people analytics, autonomous vehicle safety, robotic automation, and VLM fine-tuning — always focused on measurable impact: faster inference, higher accuracy, and real-world deployment at scale.

Core Focus Areas

Real-Time Video Analytics

Multi-stream processing with NVIDIA DeepStream, tracking, and detection at scale.

Edge AI & Optimization

TensorRT, CUDA profiling, GPU optimization for sub-50ms inference.

Face Recognition & Biometrics

InsightFace, pgvector, KNN search with production-grade accuracy.

VLM & LLM Deployment

Fine-tuning Vision-Language Models and deploying with vLLM/llama.cpp.

Experience

Where I've worked

Computer Vision Engineer

Nov 2025 — Present

Adagrad AI · Pune, India

  • Designed and developed the complete real-time people analytics ML pipeline using NVIDIA DeepStream for two-camera edge-device deployments.
  • Developed people detection, tracking, entry/exit counting, gender classification, and age estimation workflows, achieving 96% accuracy through custom MiVOLO plugin for TensorRT-optimized PyTorch inference.
  • Built FastAPI backend services to receive ML engine payloads, process analytics data, and store operational events in PostgreSQL.
  • Implemented face recognition storage and retrieval using InsightFace embeddings, pgvector-based vector databases, and KNN similarity search, reducing search latency from 19ms to 10ms.
  • Implemented asynchronous inference support, increasing real-time multi-stream processing performance from 9 FPS to 25 FPS.
  • Fine-tuned VLM model using Supervised Fine-Tuning to detect road hazards including fallen objects, fire, and smoke from visual inputs.
  • Profiled inference workloads with NVIDIA Nsight Systems and Nsight Compute, reducing GPU stall time from 1150ms to 1ms.
  • Containerized and deployed ML services and backend components using Docker across edge-device workflows.
NVIDIA DeepStreamTensorRTFastAPIpgvectorDockerVLM

Software Development Engineer

Mar 2023 — Oct 2025

Rosa Technology · Jamshedpur, India

  • Designed and developed an automated robotic rebar testing system, improving operational efficiency by 60% over manual workflows.
  • Programmed robotic arm automation for pick-and-place operations and integrated automation logic for autonomous rebar handling during testing.
  • Developed CNN-based computer vision pipelines for rebar localization, elongation measurement, breakage detection, and quality assessment.
  • Implemented client-defined formula-based decision logic and integrated ML, CV, and automation components using Python, OpenCV, and TensorFlow.
PythonOpenCVTensorFlowCNNsRobotics

Skills

Technical expertise

Languages

Python
C++

ML / DL Frameworks

PyTorch
TensorFlow
Hugging Face
vLLM
llama.cpp

Computer Vision

NVIDIA DeepStream
TensorRT
OpenCV
Object Detection
Face Recognition

Optimization & MLOps

Docker
CUDA
Triton Inference Server
ONNX Runtime
GPU Profiling

Backend & Data

FastAPI
PostgreSQL
pgvector
REST APIs

Tools

Git
Linux
Nsight Systems
MediaPipe

Projects

Featured work

Real-Time People Analytics Pipeline

End-to-end multi-camera edge analytics system using NVIDIA DeepStream with people detection, tracking, counting, gender/age classification, and face recognition.

96% Accuracy
25 FPS Throughput
1ms GPU Stall
DeepStreamTensorRTMiVOLOInsightFacepgvectorFastAPI

VLM Road Hazard Detection

Fine-tuned Vision-Language Model using SFT to detect road hazards — fallen objects, fire, smoke — from real-world visual inputs for autonomous vehicle safety.

VLM Architecture
SFT Training
VLMSFTHugging FacePyTorchTensorRT

Automated Robotic Rebar Testing

Robotic rebar testing system with CNN-based CV pipelines for localization, elongation measurement, breakage detection, and autonomous pick-and-place operations.

60% Efficiency Gain
CNN Detection
PythonOpenCVTensorFlowCNNsRoboticsAutomation

Football Analytics System

End-to-end football analytics pipeline for automated player and ball tracking from broadcast video footage using YOLOv8 and computer vision techniques.

YOLOv8 Detection
Real-time Processing
YOLOv8OpenCVK-MeansOptical FlowPython

Face Recognition & Candidate Verification

InsightFace-based candidate verification system with pgvector storage, KNN similarity search, and MediaPipe eye-gaze monitoring for driving test supervision.

10ms Search Latency
KNN Search
InsightFacepgvectorMediaPipeFastAPIPostgreSQL

Autonomous Vehicle Safety Systems

Rule-based CV algorithms for stalled vehicle detection, wrong-direction driving detection, and poor-visibility detection across blur, adverse weather, and visual degradations.

3+ Detection Modes
Real-time Inference
OpenCVDeepStreamEdge AIPythonTensorRT

Education

Academic background

B.Tech in Computer Science and Engineering

RVS College of Engineering and Technology

2019 — 2023 · CGPA: 8.5/10

Contact

Let's connect

I'm always open to discussing new projects, creative ideas, or opportunities to be part of something exciting.

tuhinacharjee91@gmail.com LinkedIn GitHub +91-7903067070