Bhuvnesh Kumar.

01. About me

A bit about me

02. Certifications

Continued learning

Recent courses I've completed and recommended internally.

• Building Deep Agents with LangGraph

  LangChain Academy

  2025
  View certificate ↗
• Search in the LLM Era for AI Engineers

  Vector Search Cohort

  2024
  View certificate ↗
• AI Agents with LangGraph & Tavily

  DeepLearning.AI

  2024
  View certificate ↗

02. Current Work

What I'm shipping right now

Flagship initiatives I'm leading at VCTI for Cisco optical networking.

03. Skills

Technologies & Tools

The stack I reach for when building agents, vision systems, and the infrastructure that runs them in production.

04. Career Changelog

My career as a changelog

Each role is a release, each win an Added entry.

1. 4.0.0 Technical Team Lead — GenAI & Computer Vision

   @ VCTI (Client: Cisco) · Bangalore, India · Nov 2022 → Present

   Leading AI agent platforms for Cisco optical networking — shipping the first MCP-protocol AI tool server in CONC 26.1.1, the AINetOps (C-ONIDAA) product line, and a suite of Generative-UI dashboards. Owning architecture, security, delivery, and team enablement across 10+ engineers on parallel workstreams.

   • Python
   • LangGraph
   • LangChain
   • MCP
   • FastAPI
   • React
   • Docker
   • Kubernetes
   • AWS EKS
   • mem0
   • Docling
   • EFK

   ### Added

   • + CONC MCP Server (26.1.1): Led delivery of the first MCP-protocol AI tool server for Cisco optical networking — positioned as a standard microservice in the NxF framework.
   • + AINetOps (C-ONIDAA) Beta: Owned end-to-end development and early field-trial delivery (FastAPI, LangGraph, React, MCP, Docker).
   • + Deep Agent for Automated RCA: Architected a LangGraph workflow that autonomously investigates customer incidents — traverses logs and process flows, generates L1 triage reports, and reduces MTTR by 50%.
   • + Agentic documentation pipeline: AutoGen + LangGraph system that turns meeting transcripts into technical docs, reducing manual effort by 90%.
   • + Production RAG: LlamaIndex contextual RAG over private enterprise data hitting 80% context recall via custom embedding fine-tuning.
   • + Self-hosted Generative UI: Interactive charts/tables in chat without external LangSmith dependency — a first in the CONC ecosystem. POC → production dashboards (NHot, Triage, Alarm, Utilization) in 8 weeks.
   • + Manager's Triage Dashboard: Generative-UI dashboard collating critical data sources with interactive deep-dive views.
   • + NHot: Course-generation platform for new Cisco interns and internal employees.
   • + Open-source optical encoder: Guided team to publish the first open-source encoder for optical networking on Hugging Face.
   • + Cisco Optical AI Agent + CONC MCP client: Built and delivered for the PONC demo.
   • + Cross-platform agent integration: Atlassian, ELK, and CDETS tools wired into agents for cross-platform intelligence.
   • + Persistent agent memory: Established mem0-based recall across sessions supporting multi-turn Q&A for complex troubleshooting.
   • + Document conversion: Docling-based pipelines and a doc-to-doc agent for higher structural fidelity.
   • + Reusable LLM tooling patterns: FunctionTool wrappers, tool cataloging, consistent descriptions.
   • + MLOps: Migrated inference pipelines to AWS EKS; GPU optimization reduced inference latency and cloud costs; standardized Docker workflows for CI/CD; EFK stack for observability.
   • + Computer Vision: Improved pole-detection accuracy by 48% using DeepLabv3+ semantic segmentation on satellite imagery.
   • + Security-first culture: API lockdown, Nginx rate limiting, JWT auth, OWASP review, DDoS protection, CSDL compliance — completed before any external deployment.
   • + OFC Demo: Made AI products demo-ready and deployed for Cisco's international OFC showcase.
   • + Team enablement: Trained the AI team on LangGraph, MCP, LangChain, and agent fundamentals; scaled output across 10+ engineers on parallel workstreams (CSDL, dashboards, token optimization, DeepEvals, Deep Agents, SSO).
   • + Cutting-edge adoption: Continuously evaluated and incorporated MCP protocol, Generative UI, Deep Agents, and DeepEvals into production.
   • + Feature POCs delivered: Gen UI, SMX evaluation, SSO, multi-agent, RCA deep agents, MCP orchestration.
   • + AI pole measurement methodology: Decomposed the problem and provided optimal solutions for each part.
   • + In-house GPU infrastructure: Set up GPU machine with full AI stack for the team.

   ### Changed

   • ~ VCTI ↔ Cisco alignment across Core Platforms, COSM, CONP, and NCS2K legacy; built RCA agent for dev productivity and nightly issue resolution.
2. 3.0.0 Senior Software Engineer — Computer Vision

   @ Wipro (Client: HSBC Bank) · Gurgaon, India · Oct 2021 → Nov 2022

   Built production computer-vision and document-intelligence systems for HSBC's KYC and document-processing pipelines.

   • Python
   • YOLOv5
   • VGG16
   • TrOCR
   • PaddleOCR
   • TF-IDF

   ### Added

   • + High-precision object detection: ID-card localization with YOLOv5 — 95% mAP in production.
   • + Multimodal document understanding: Hybrid classifier combining VGG16 visual features with TF-IDF text features — boosted document classification accuracy from 70% → 95%.
   • + Inference optimization: Deep-profiled Python pipelines and refactored bottlenecks for a 4× reduction in request latency and significant memory savings.
   • + OCR solutions: Deployed Transformer-based OCR (TrOCR, PaddleOCR) for multi-lingual handwritten documents at 90% accuracy.
3. 2.0.0 Senior Software Engineer — Computer Vision

   @ ARTIVATIC.AI · Bangalore, India · Jun 2021 → Oct 2022

   Built computer-vision pipelines for insurance — e-KYC, identity verification, predictive risk attributes from facial imagery, and signature fraud detection.

   • Python
   • OpenCV
   • dlib
   • Siamese Networks
   • ResNet50V2
   • YOLOv4

   ### Added

   • + Identity verification pipeline: Complete e-KYC using Siamese Networks for face matching and dlib for liveness detection — critical for insurance fraud prevention.
   • + Predictive modelling: Fine-tuned classifiers to predict Age, Gender, BMI, and Smoker Status from facial imagery.
   • + Document intelligence: Improved signature fraud detection accuracy from 45% → 81% using a ResNet50V2 backbone for signature matching.
4. 1.0.0 Full-Stack Developer

   @ Snap-on Business Solutions · Noida, India · Jun 2018 → Jun 2021

   First professional role. Built scalable web applications and an automated chassis-number detection system using classic image-processing techniques.

   • Angular
   • Java
   • OpenCV
   • Agile

   ### Added

   • + Chassis-number detection: Automated system using image processing and OpenCV.
   • + Web applications: Scalable apps built with Angular and Java, following Agile methodologies and MVC design patterns.

05. Education

Where I learned the fundamentals

• B.Tech in Computer Science & Engineering
  Uttarakhand Technical University, Dehradun
  2012 – 2016 · 71%
• Senior Secondary (CBSE)
  DAV Centenary Public School, Haridwar
  2012

07. Projects

Things I've built

A selection of personal projects — Colab and Kaggle notebooks I've shipped over the years. For professional work, see the Career Changelog.

Car Key-Points Detection with MobileNetV1

Feb 2020

Predicts windshield and headlight key-points on car images using a fine-tuned MobileNetV1 backbone with custom regression heads, trained on a tiny dataset (90 images) heavily augmented to 9,000 samples.

• Python
• TensorFlow
• MobileNetV1
• Keras
• Augmentation

Read case study → Open in Colab ↗ GitHub ↗

Bank Loan Customer Identification

Classification model predicting which retail-banking customers are likely to accept a personal-loan offer — a classic targeted-marketing use case.

• Python
• scikit-learn
• Classification
• EDA

View on Kaggle ↗

High-Performance Concrete Strength Prediction

Regression model predicting compressive strength of concrete mixes from cement, aggregates, water, and admixture composition.

• Python
• scikit-learn
• Regression
• Feature Engineering

View on Kaggle ↗

COVID-19 Mask Detection

Real-time detector that identifies people wearing or not wearing face masks — built during the COVID-19 pandemic as a public-safety screening aid.

• Python
• OpenCV
• CNN
• MobileNet

Open in Colab ↗

Face Recognition with VGG

Identity-classification model built on a pretrained VGG backbone with custom head layers and embedding-based matching.

• Python
• TensorFlow
• VGG
• Transfer Learning

Open in Colab ↗

Multi-Person Face Segmentation with U-Net

Pixel-accurate face segmentation across multiple people per image using a U-Net encoder-decoder trained on the WIDER FACE dataset.

• Python
• TensorFlow
• U-Net
• Semantic Segmentation

View on Kaggle ↗

IMDB Sentiment Classification

Binary sentiment classifier for IMDB movie reviews using embeddings + recurrent layers, with comparisons against classical bag-of-words baselines.

• Python
• Keras
• NLP
• LSTM

Open in Colab ↗

Parkinson's Diagnosis with Random Forests

Ensemble classifier diagnosing Parkinson's disease from voice-measurement features — feature importance analysis exposes the most discriminative bio-signals.

• Python
• scikit-learn
• Random Forest
• Ensemble

View on Kaggle ↗

Pneumonia (Lung Opacity) Detection

CNN-based detector that flags lung opacity regions in chest X-rays — a screening aid for radiologists.

• Python
• TensorFlow
• CNN
• Medical Imaging

Open in Colab ↗

Raccoon Detection with Bounding Boxes

Object-detection model that localizes raccoons in natural images with predicted bounding boxes.

• Python
• TensorFlow
• Object Detection

Open in Colab ↗

Salt vs Sediment Segmentation (TGS Salt Challenge)

Semantic segmentation of subsurface salt deposits from seismic imagery — Kaggle's TGS Salt Identification Challenge.

• Python
• Keras
• U-Net
• Semantic Segmentation

Open in Colab ↗

Sarcasm Detection in News Headlines

Trained on the Sarcasm Headlines dataset — distinguishes sarcastic from genuine news headlines using embeddings and convolutional/recurrent text models.

• Python
• Keras
• NLP
• Embeddings

Open in Colab ↗

Ship vs Iceberg Classification (Siamese)

Siamese-network classifier that distinguishes ships from icebergs in dual-polarization satellite radar imagery — Statoil/C-CORE challenge.

• Python
• Keras
• Siamese Networks
• Satellite Imagery

Open in Colab ↗

Vehicle Silhouette Classification

Multi-class classifier identifying vehicle types (bus, van, car) from extracted silhouette features — Statlog Vehicle Silhouettes dataset.

• Python
• scikit-learn
• Classification
• Feature Selection

View on Kaggle ↗