Satyam Sahoo

AI Researcher

Bhubaneswar, Odisha, India2 yrs 3 mos experience

Key Highlights

Gold medalist at Inter IIT Tech Meet 12.0
Built advanced AI systems for voice and image processing
Led robotics initiatives at IIT Patna

Stackforce AI infers this person is a Machine Learning and Robotics specialist with a focus on AI-driven solutions.

Contact

Skills

Core Skills

Natural Language Processing (nlp)MlopsAudio EngineeringMachine LearningGraph NetworksRoboticsInternet Of Things (iot)Computer VisionEngineering

Other Skills

Go (Programming Language)Amazon Web Services (AWS)kafka-toolsWebSocketgRPCFastAPIDockerDeep LearningPattern RecognitionRobot Operating System (ROS)Visual SLAMReinforcement LearningSOLIDWORKSESP32 MicrocontrollersArduino IDE

About

Obsessed.

Experience

2 yrs 3 mos

Total Experience

10 mos

Average Tenure

Current Experience

Eternal

GenAI Intern

May 2025 – Oct 2025 · 5 mos · Gurugram · On-site

● Built a production-grade Gemini Live–based Speech-to-Speech (S2S) pipeline
Designed and deployed a low-latency S2S architecture enabling natural, empathetic voice interactions with tool calling. This system now handles 40%+ of internal weekly calls, powering multiple conversational workflows across the company.
● Engineered a real-time audio microservice from scratch
Developed a denoising and signal-processing microservice with:
Silero-VAD integration for voice activity detection
A 20 ms WebSocket server–client stack in Go for streaming audio
ONNX-based inference for efficient model deployment
FFT-based spectral subtraction for real-time noise removal
Dead-air mitigation using PCM16-based timers
Also implemented real-time tool calling with gRPC + Kafka, supported by a Redis-cached transcript classifier to ensure deterministic fallback behavior.
● Created an automated Voicebot Evaluation Pipeline
Developed an end-to-end evaluation system for assessing voice interactions at scale. It used ETL jobs + HTTPS calls to GPT-4o-audio to evaluate prompt adherence, latency, and quality. Designed an asynchronous evaluation framework to bypass rate limits and handle concurrent multi-call evaluations reliably.
● Built and deployed a high-throughput Super-Resolution Pipeline
Productionized a Real-ESRGAN–based enhancement pipeline processing ~30% of weekly image volume.
Optimizations included:
Custom preprocessing reducing inference latency from 30+ sec → ~10 sec
FastAPI + Docker deployment for horizontal scaling
S3 integration for storage and large-file workflows
Added blur detection for quality scoring
Built a border detection model that reduced false positives by 63%

Go (Programming Language)Natural Language Processing (NLP)Amazon Web Services (AWS)kafka-toolsMLOpsWebSocket+3

Siemens digital industries software

ML Intern

Dec 2024 – May 2025 · 5 mos · Remote

Implemented Hierarchical CADNet, a graph neural network, for machining feature recognition in CAD models using
hierarchical B-Rep graphs to extract geometric and topological information.
Achieved state-of-the-art segmentation accuracy of 97.3% on the MFCAD dataset and 95.37% on the MFCAD++
dataset, leveraging graph convolution neural network, attention based weighted node features through graph
attention neural network, and graph-to-graph knowledge transfer for robust feature recognition
Built Graph AutoEncoder for unsupervised feature extraction and Cad Body Classifier

Graph NetworksDeep Learning

Infivr

ML/AI Intern

Aug 2024 – Oct 2024 · 2 mos · Remote

Fall Detection System using 3D Point Clouds from Sparse RF Signals
Developed and implemented an advanced fall detection system by reconstructing 3D human point clouds from sparse RF signal data, leveraging Temporal Convolutional Networks (TCNs), SMPL fitting, and Autoencoders (AEs).
Temporal Convolutional Networks (TCNs): Modeled the temporal sequences of sparse point cloud data to capture human movement patterns over time. By analyzing sequential 3D point clouds, TCNs learned temporal dependencies, enabling accurate detection of abrupt posture changes associated with falls.
Autoencoders (AEs): Utilized Autoencoders to reconstruct complete human body structures from sparse point cloud inputs. The AE architecture compressed sparse input data into a low-dimensional latent space, then reconstructed the full 3D point cloud, preserving critical structural features necessary for interpreting human posture and motion dynamics.
Point Cloud to SMPL Fitting: Fitted the SMPL model to point cloud data without explicit joint information using iterative optimization techniques. This process minimized the distance between the point cloud and the SMPL model's vertices, resulting in an accurate representation of the human body.
This system effectively integrated TCNs for temporal sequence analysis with AEs for 3D reconstruction, creating a robust solution for fall detection from sparse RF-based point clouds. By combining both spatial and temporal information, this approach significantly enhanced the accuracy of fall detection.

Deep LearningPattern RecognitionMachine LearningComputer Vision