Everything I’ve built, grouped by where I built it. Selected code and open tools will be published at github.com/phvpavankumar soon.
Kearney · 2026 – present
Agentic analytics platform for enterprise finance & supply chain
An AI-powered analytics platform where executives ask business questions in plain language and receive governed KPI cards, charts, tables, narrative summaries and suggested follow-ups — alongside executive dashboards covering finance, commercial, operations, receivables and supply-chain performance. The platform spans 27 registered analytical tools and over 100 API endpoints across areas including product contribution and margin, cost-to-serve, quality cost, cash conversion cycle, AR/AP and payment behaviour, inventory and DIO, OEE, case fill and OTIF.
My work: designing and enhancing analytical tools; LangGraph-based AI orchestration with intent classification and tool routing; SAP Datasphere/OData integration; establishing source-to-UI KPI lineage and reconciliation practice — tracing discrepancies through extraction, transformation, business mappings, AI routing and presentation; automated SAP reconciliation scripts, regression tests and fiscal-period controls; and production trace analysis with Langfuse. Stack: Python, FastAPI, LangGraph, React, PostgreSQL, Redis, Docker, Azure Kubernetes Service, Microsoft Entra.
Process intelligence from video
A three-service system that turns recorded business-process videos into structured outputs: step-by-step documentation, chapters, HTML guides, BPMN process diagrams, test cases, and a searchable knowledge repository. Uploads become asynchronous jobs claimed by AI workers through a database-driven control plane with statuses, claims and heartbeats — no separate queueing platform required.
My work centred on two areas. First, test-case generation and delivery — backend/AI integration and the download APIs that expose generated artifacts. Second, the knowledge repository layer that turned the platform from “video in, document out” into “video in, reusable organisational knowledge out”: auto-ingest of generated documentation, semantic chunking and embedding with pgvector, hybrid retrieval combining vector search with governed SQL, evidence merging and answer synthesis, task-scoped chat history, and production hardening — SQL parameterisation, ingest guardrails, failure metrics and runtime decoupling. Stack: Python, FastAPI, APScheduler, PostgreSQL + pgvector, Azure Blob, Nuxt 3/Vue 3.
Solum · 2024 – 2026
Multi-camera retail shelf-vision system
Real-time product recognition and out-of-stock analytics across retail shelf environments, built on YOLOv7/YOLOv9 with TensorRT acceleration and LightGlue feature matching, with high-precision multi-camera calibration. Delivered sub-200 ms response times and 15–20 FPS inference by refining sensor-driven session flows and GPU-optimised execution, with end-to-end CI/CD for model deployment.
RFID ML-serving platform
High-performance ML-serving APIs for real-time shelf mapping, built on Flask and Gunicorn and integrated with Azure Event Hubs and MongoDB. Engineered hot-reload capability so models update without downtime, plus monitoring, structured logging, automated scalability and fully containerised deployment.
LLM-powered log intelligence
A framework that transforms unstructured production logs into structured insight for root-cause analysis, combining fuzzy clustering, ReAct-based LangChain agents and locally hosted LLaMA models. Brought log investigation down from hours to seconds and streamlined over a thousand review cycles through SME automation.
Sony · 2023
Edge AI on IMX500 smart cameras
Custom object-detection and segmentation models built and optimised for Sony’s embedded devices, tuned for on-device speed and accuracy, with real-time detection deployed on IMX500 intelligent vision sensors. Transfer-learning workflows cut training time and integration effort substantially.
Ignitarium · 2021 – 2024
Neural networks on microcontrollers
Led architectural refinement of the FastestDet object-detection network for ultra-lightweight deployment — a 35% accuracy improvement — and shipped real-time detection models on Renesas MCUs, where there is no cloud to fall back on. Built the surrounding model-development pipelines: optimisation, training, evaluation and deployment readiness.
RGB-D reconstruction and binocular vision
3D reconstruction and stereo-vision pipelines enabling precise alignment and robust defect detection for industrial inspection, spanning image registration, semantic segmentation and 3D object reconstruction.
Domain-specific LLM Q&A and GenAI systems
Domain-specific question-answering systems and custom neural networks accelerating early defect detection across railway, telecom and solar industries. Alongside these, end-to-end GenAI builds including Stable Diffusion–based text-to-image models, LLM-powered analytics pipelines, and vector-database retrieval systems using LangChain with custom prompt-engineering frameworks — plus fine-tuning work across Falcon, LLaMA-2 and BERT-family models with PEFT.