Shiv Kumar | Lead Engineer – GenAI & Agentic AI

BEV-IMTP: Lightweight Multi-Task Learning for Joint Bird's-Eye-View Instance Mapping and Trajectory Prediction

Sun, 16 Aug 2026 00:00:00 +0000

Publication Details

Journal: Discover Artificial Intelligence
Publisher: Springer Nature
Published online: 16 August 2026
Volume and issue: 6(1)
Article: 811
DOI:

IFRS Agents — Invoice and FP&A AI (Google ADK)

Mon, 01 Sep 2025 00:00:00 +0000

Project Details

Agent Architecture

Built Gemini/Google ADK agents deployed on Vertex AI for finance automation.

Invoice Intelligence

Developed invoice risk detection using vendor history, confidence scoring, and evidence-backed signals.

Automated Routing

Enabled automated severity-based routing for faster investigation and escalation.

FP&A Analysis

Compared SAP actuals with budget, forecast, and prior-period data, with validation checks for data quality.

Executive Reporting

Identified root-cause drivers and generated management-ready variance commentary and exception reports.

Production Operations

Deployed with Vertex AI, Agent Engine, PostgreSQL/Cloud SQL, Prometheus observability, and Cloud Run.

Tech Stack

AI & Platform

Python GCP GenAI LLMs Agents Google ADK Vertex AI Agent Engine FastAPI

Data & Operations

Cloud SQL PostgreSQL Prometheus Docker Cloud Run

AACA — AI Contact-Center Platform

Fri, 01 Aug 2025 00:00:00 +0000

Project Details

Microservices Architecture

Separated a latency-sensitive realtime voice edge from a stateful business control plane for independent scaling, deployment, and ownership.

Low-Latency Voice Edge

Developed FastAPI and WebSocket services for streaming call audio and realtime agent interactions.

Telephony Integration

Bridged Twilio Media Streams to the OpenAI Realtime API with native G.711 u-law audio to minimize transcoding overhead.

Speech Performance

Delivered bidirectional speech-to-speech interaction with sub-second voice AI latency.

RAG & Tool Calling

Used OpenAI embeddings and Milvus retrieval with tools for knowledge search, customer verification, and ticket workflows.

Secure Operations

Enforced identity verification before sensitive tool operations inside live voice sessions.

Agent Handoff

Designed Redis pub/sub channels for cross-instance WebSocket delivery, call assignment, transfer, and human handoff coordination.

Deployment Stack

Containerized MongoDB, Redis, Milvus, MinIO, and PostgreSQL with Docker Compose for Cloud Run horizontal scaling.

Tech Stack

Voice & AI

Python GCP GenAI LLMs Agents RAG OpenAI Realtime API Twilio WebSockets

Backend & Data

FastAPI Milvus Redis MongoDB JWT

Deployment

Docker Docker Compose MinIO Postgres Cloud Run

SEBI Search — Regulatory RAG Platform

Sun, 01 Jun 2025 00:00:00 +0000

Project Details

Document Pipeline

Built a full RAG platform over thousands of SEBI legal and regulatory documents.

PDF Understanding

Parsed PDF files with Docling and created heading-aware document chunks.

Semantic Retrieval

Generated Jina v5 embeddings and stored them in Milvus for vector search.

Hybrid Search

Combined Jina semantic retrieval with BM25 keyword search for stronger recall.

Reranking Quality

Added ModernBERT cross-encoder reranking, reaching an 80% retrieval hit rate and 75.8% nDCG.

Grounded Generation

Implemented citation-constrained generation with vLLM/OpenAI, grounding answers to exact file, page, and order metadata.

Production Reliability

Added Redis and FAISS caching, Prometheus metrics, health probes, concurrency controls, and request timeouts.

Offline Deployment

Delivered 100% offline Docker Compose deployment for restricted SEBI environments on RHEL 9 air-gapped VMs.

Tech Stack

Retrieval & NLP

Docling Jina v5 Embeddings BM25 ModernBERT Reranker vLLM OpenAI APIs Hybrid Search

Platform & Operations

Python FastAPI Milvus Redis FAISS Prometheus Docker RHEL 9

ClothFormer: A Boundary-Aware Self-Attention Network for Human Outfit Parsing

Wed, 01 Nov 2023 00:00:00 +0000

Read the paper on .

DCSP-RCNN: A Network for Person Detection with Color, Size and Pattern Characteristic Parsing

Wed, 01 Nov 2023 00:00:00 +0000

Read the paper on .

Experience

Tue, 24 Oct 2023 00:00:00 +0000

Image Stitching — Qualcomm Edge Deployment

Sun, 01 Jan 2023 00:00:00 +0000

Project Details

Model Conversion

Transformed a feature-detection deep learning model into ONNX and Qualcomm target-device formats.

C++ Integration

Created a C++ pipeline combining deep learning feature detection and matching.

Realtime Deployment

Deployed the pipeline on Qualcomm hardware for realtime edge processing.

Performance Gain

Reduced CPU runtime by 75% through edge pipeline optimization.

Tech Stack

ML & Debugging

Python NumPy PyTorch OOP PDB Debugger

Edge & Systems

C++ ONNX QNN VS Code Linux

CES — Bird-Eye-View Instance Segmentation

Sat, 01 Jan 2022 00:00:00 +0000

Project Details

Instance Segmentation

Trained SOLOv2 on customized datasets for automotive vision workloads.

Bird's-Eye View

Implemented OpenCV image blending to improve visual outputs of BEV images.

Multi-Model Design

Transformed instance segmentation into a multi-model architecture with multi-image classification.

ML Pipelines

Developed and executed end-to-end training and testing pipelines.

Model Explainability

Used LIME for YOLO feature extraction and visualization.

Tech Stack

Computer Vision

Python NumPy PyTorch SOLOv2 mmdet YOLO LIME

Engineering Tools

OOP PDB Debugger OpenCV VS Code Linux

Deep Learning Kernels — Accelerator Optimization

Fri, 01 Jan 2021 00:00:00 +0000

Project Details

Accelerator Framework

Developed a PyTorch and TensorFlow software suite for high-performance accelerator training and inference.

Kernel Debugging

Analyzed and resolved kernel bugs and performance bottlenecks.

TPC Mapping

Mapped PyTorch kernels to TPC kernels in C++.

Custom Operators

Developed operators using combinations of TPC kernels.

Model Export

Converted deep learning models to ONNX for deployment.

Graph Validation

Used Netron to visualize and validate exported model graphs.

Performance Focus

Optimized execution across accelerator training and inference workloads.

Tech Stack

Frameworks & Data

Python NumPy Pandas PyTorch TensorFlow

Acceleration & Tooling

C++ TPC Kernels ONNX Netron VS Code Linux Git JIRA

AgVerse — Product Analytics ML

Wed, 01 Jan 2020 00:00:00 +0000

Project Details

Telemetry Platform

Brought up Firebase server support to pull and analyze telemetry data.

Data Processing

Created an on-demand pipeline mechanism for integrating preprocessing modules.

Analytics Workflow

Processed telemetry data with Python, Pandas, and Firebase services.

HMM Estimation

Performed Baum-Welch algorithm-based parameter estimation.

State Modeling

Performed hidden Markov model parameter and state estimation.

Tech Stack

Data & ML

Python Pandas HMM

Platform

Firebase