Kishor Kumar Paroi
Photograph of Kishor Kumar Paroi
Kishor Kumar Paroi
Principal AI Systems Architect & Founder @ KoomAI

πŸ“„ 1-Page CV πŸ“‘ 2-Page ATS πŸ’Ž Pitch Deck β†— 🎬 $1.4M Reel β†—

I am an AI systems architect and engineer working at the intersection of distributed ML infrastructure, sub-millisecond retrieval, and production agent mesh networks. Currently, I am the founder and lead architect at KoomAI, where I built an enterprise WebRTC high-performance video communication engine sustaining <120ms glass-to-glass latency across 50,000+ recording minutes.

Previously, I served as a Forward Deployed AI Engineer at phi0.ai in Europe, architecting production LangGraph agent workflows (99.4% task completion across enterprise workloads) and high-throughput Hybrid RAG engines (1) that curtailed p95 latency to <35ms and saved $18K/month in compute expenses. Earlier, at Samsung AI Core / Care Health, I engineered distributed data pipelines across AWS EMR and accelerated 8B/70B LLM training and inference on NVIDIA H100 clusters with FlashAttention-2, custom fused CUDA kernels, and PyTorch FSDP / DeepSpeed ZeRO-3. I hold a B.Sc. in Computer Science & Engineering from RUET (ranked 483rd nationwide in Bangladesh).

My work spans low-latency systems and mathematical foundations: from authoring Rust SIMD vector search accelerators (<0.5ms scoring latency, 50x speedup over standard implementations) and Go microservices handling high concurrency, to building zero-trust enterprise guardrails (BotShield WAF, NVIDIA NeMo) blocking 95%+ of prompt injections and jailbreaks.

When not designing production lakehouses or fine-tuning models with LoRA/DPO, I write experimental systems software ranging from a Git storage engine written in pure Rust, to an offline-first industrial IoT edge telemetry engine with SQLite write-ahead caching, to a real-time multimodal receipt arbitration engine using DHash perceptual hashing and G-Eval.

Below is a curated collection of my systems & architectures, active open-source code, work history, and technical learning curriculum. You can also inspect the verified credentials: πŸ“„ 1-Page Visual Resume (PDF) • πŸ“‘ 2-Page ATS Resume (PDF).

Digital Ecosystem: πŸ’Ž $10K Executive Pitch Deck • pptx.kishor.tech β†— 🎬 $1.4M Architecture Reel • YouTube β†— πŸŽ“ Kishor.tech • Main Academic Portfolio πŸ’» macos-portfolio • macOS Desktop Portfolio 🧊 3d • Interactive 3D WebGL Platform ⚑ hacker • 3D Cyber Hacker Workstation β–Ά koom.kishor.tech • Sub-120ms Video Intelligence πŸ“… Book 15m Zoom • Calendly Instant Sync
Founder Experience & Startup Platform: KoomAI

KoomAI β€” Enterprise Asynchronous Video Intelligence Platform • Founder & Principal Architect
Orchestrating Desktop (Electron), Chrome Extension (Manifest V3), Backend (Express/AWS), and Full-Stack Next.js 16 App Router. [Live Platform • GitHub]

koom.kishor.tech • WebRTC SVC Engine
// Sub-120ms Glass-to-Glass WebRTC Pipeline
const peer = new RTCPeerConnection({ sdpSemantics: 'unified-plan' });
peer.addTransceiver('video', { direction: 'sendrecv', streams: [mediaStream] });
const stats = await peer.getStats();
// p99 latency: 114ms • 60 FPS 1080p • 0% packet stutter
LIVE STARTUP 50K+ Video Minutes • $24K MRR

KoomAI β€” Sub-120ms Real-Time Desktop, Extension & Web Platform
I founded KoomAI to challenge traditional bloated screen-recording tools. By implementing an adaptive WebRTC pipeline with Scalable Video Coding (SVC), hardware-accelerated NVENC/VideoToolbox native capture in Electron, and offscreen Manifest V3 browser capture, we achieved 60 FPS 1080p video recording with zero packet stutter across 50,000+ recording minutes. The Next.js 16 cockpit with AWS S3 multipart streaming cut video storage expenses by 62%.

Integrated an autonomous 20-agent LangGraph copilot mesh (<35ms p99) and interactive video chatbot generating 1-click GitHub PRs, Jira BDD stories, and Playwright E2E tests over Whisper transcripts.

Personal NotebookLLM Studio & AI Job Recon: Built conversational two-host audio podcast generator, Principal Architect system design drills (Kafka, CDC, Feast, Saga), and automated ai-job-search Trojan Horse proposals with tailored ATS CVs, candidate briefing dossiers, and automated outreach pipelines.

Pomodoro Chrome MV3 & Client Invoicing: Engineered screen-recording extension with integrated Pomodoro focus sprints, tab usage monitor, and 432Hz meditation audio, coupled with instant Stripe client invoicing ($24K MRR pilot).

Technical Projects & Architectures

Engineered with strict enterprise rigor: Latency vs Cost vs Accuracy vs VPC Data Privacy. Click any diagram thumbnail to view full resolution or click titles to expand architectural benchmarks.

Zomato AI Lakehouse Architecture Diagram
πŸ” Architecture Blueprint
Code Blueprint Docs


Kishor Kumar Paroi
Enterprise Data & AI Platform Engineering • Production Lakehouse 35M+ Orders (2.3 GB) • $1.2M Saved

Enterprise Medallion Lakehouse on Snowflake and PySpark Delta Lake processing 35M+ orders using dbt Core SCD-2, reducing compute expenditure by 42% ($18K/mo) with sub-5ms feature serving via Feast over Redis.

Architectural Abstract: Devised streaming feature ingestion via Apache Airflow, Debezium CDC WAL capture, and Feast Feature Store over Redis, bringing feature serving latencies below 5ms. Built an automated multimodal dispute resolution engine using DHash perceptual hashing and G-Eval chain-of-thought verification to neutralize delivery fraud claims by 68%, safeguarding $1.2M in annual partner margins.

Key Stack: Snowflake, PySpark, Delta Lake, dbt Core, Debezium CDC, Feast, Kafka, OpenCV, Presidio PII, AWS S3 Glacier.

Runbook: curl -fsSL https://raw.githubusercontent.com/KishorKumarParoi/zomato-analysis/main/bootstrap.sh | bash

Hybrid Anime LLMOps Architecture
πŸ” Architecture Blueprint
Code Blueprint Docs


Kishor Kumar Paroi
Distributed LLMOps • High-Throughput Recommender Mesh Sub-0.5ms SIMD (50x Speedup) • NDCG@10 = 0.94

Polyglot service mesh integrating a high-concurrency Go API Gateway, low-latency Rust SIMD AVX-512 vector scoring engine (<0.5ms), and dynamic Multi-LoRA multiplexing on vLLM cutting GPU spend by 68%.

Architectural Abstract: Curtailed ranking latency from 25ms to <0.5ms (a 50x speedup), diminishing CPU costs by 64%. Synthesized Dual-Tower Neural Collaborative Filtering (NCF), ChromaDB HNSW dense vectors, and BM25 sparse search using Reciprocal Rank Fusion to achieve NDCG@10 = 0.94. Served LoRA adapters via dynamic vLLM multiplexing on single-GPU instances, elevating throughput by 3.4x while orchestrating multi-cloud GitOps delivery with ArgoCD across AWS EKS, GCP GKE, and Azure AKS.

Key Stack: Rust SIMD, Go Gateway, vLLM Multi-LoRA, PyTorch, Kubeflow v2, MLflow, ArgoCD, Kafka, KEDA.

Runbook: curl -fsSL https://raw.githubusercontent.com/KishorKumarParoi/Hybrid-Anime-Recommender/main/bootstrap.sh | bash

KoomAI WebRTC Platform Architecture
πŸ” Architecture Blueprint
Live Platform Code Blueprint


Kishor Kumar Paroi
Real-Time WebRTC SVC Engine • Native Electron C++ IPC • Next.js 16 50K+ Video Mins • <120ms Glass-to-Glass • $24K MRR

Full-stack Loom alternative engineered with WebRTC adaptive bitrate streaming (SVC), native Electron desktop client with OS-level audio mixing, Manifest V3 Chrome extension, and autonomous 20-agent LangGraph video copilot.

Architectural Abstract: Built native Electron desktop client with OS-level audio mixing and NVENC/VideoToolbox hardware acceleration, achieving 45% lower idle CPU utilization. Engineered Next.js 16 App Router cockpit with AWS S3 multipart parallel uploads and CloudFront edge delivery. Integrated an autonomous 20-agent LangGraph copilot mesh (<35ms p99) and interactive video chatbot generating 1-click GitHub PRs, Jira BDD stories, and Playwright E2E tests over Whisper transcripts.

Key Stack: WebRTC, Electron C++, Next.js 16, Express CDN, AWS S3, CloudFront, Stripe, LangGraph, Whisper.

Runbook: git clone https://github.com/KishorKumarParoi/Koom-AI.git && cd Koom-AI && npm install && npm run dev

macOS Sequoia Desktop OS Architecture & UI
πŸ” OS Architecture Blueprint
Live macOS OS Code Blueprint


Kishor Kumar Paroi
Interactive OS Architecture • Desktop Simulation • React 19 • Web Audio API Sub-50ms Dock Bounce • Multi-Window Z-Index • Live Telemetry

Full-fidelity desktop operating system simulation of macOS Sequoia in the browser, featuring draggable window manager, interactive Finder, Spotlight search, reactive Unix terminal emulator, audio synthesizer, and edge-synced visitor telemetry.

Architectural Abstract: Built native desktop-grade windowing subsystem supporting draggable, resizable, and minimizable windows with collision detection and Z-index layering. Implemented interactive macOS dock with magnification physics, menu bar system status widgets, Spotlight fuzzy search, full-blown reactive Unix shell emulator with standard terminal utilities, Web Audio API sound engine, and Vercel Edge Serverless visitor telemetry synchronization with sub-15ms regional latency.

Key Stack: React 19, TypeScript, Tailwind CSS, Web Audio API, Lucide Icons, Vercel Edge Serverless.

Runbook: git clone https://github.com/KishorKumarParoi/macos-portfolio.git && cd macos-portfolio && npm install && npm run dev

3D Cyber Hacker Cockpit WebGL Scene
πŸ” WebGL Cockpit Blueprint
Live Hacker Cockpit Code Blueprint


Kishor Kumar Paroi
3D WebGL Metaverse • React Three Fiber • GSAP 3D Camera • Web Audio API 60 FPS WebGL • Sub-80ms First Paint • Cyber Hacker Cockpit

Futuristic 3D Cyber Hacker command center featuring 360-degree scene exploration with OrbitControls, interactive hacker room with animated CRT monitors, cyber audio synthesis, and HUD intelligence consoles with 2-page ATS resume dossier integration.

Architectural Abstract: Developed high-performance 3D scene graph with React Three Fiber, Drei, and Three.js featuring Draco GLTF compression, dynamic lighting, bloom post-processing, and responsive camera trajectories orchestrated via GSAP. Integrated terminal command line interface with dossier intelligence viewer, custom soundscape engine, and sub-100ms Vercel Edge CDN delivery.

Key Stack: React Three Fiber, Three.js, React 19, GSAP, Tailwind CSS, Web Audio API, Vite.

Runbook: git clone https://github.com/KishorKumarParoi/revive-yourself.git && cd revive-yourself && npm install && npm run dev

AI Security Guardrails Architecture
πŸ” Architecture Blueprint
Code Blueprint Docs


Kishor Kumar Paroi
AI Safety & Adversarial Defense • Production Agent Mesh 95%+ Attacks Blocked • Sub-35ms p95 • $8.4K/mo Saved

Zero-trust security defense proxy combining BotShield WAF, Portkey API gateway, and NVIDIA NeMo Guardrails, defending against 95%+ of prompt injections and jailbreaks certified via Microsoft PyRIT 0.14.0.

Architectural Abstract: Formulated a cyclic state machine in LangGraph with state checkpointers, saving $8.4K/month via <50ms semantic similarity caching. Standardized tool invocation via FastMCP with human-in-the-loop break-glass gates, and coupled two-stage retrieval (Jina v3 1024-dim dense vectors, BM25, and local ONNX FlashRank cross-encoders) to realize 96.4% context precision.

Key Stack: LangGraph, FastMCP, NeMo Guardrails, FlashRank ONNX, Portkey, Microsoft PyRIT, Qdrant, Docker.

Runbook: curl -fsSL https://raw.githubusercontent.com/KishorKumarParoi/AI_Security_Production_App/main/bootstrap.sh | bash

Smart Manufacturing Architecture Master
πŸ” Architecture Blueprint
Code Blueprint Docs


Kishor Kumar Paroi
Edge AI • Industrial Computer Vision • Kubernetes 31% Defect Reduction • $420K Saved

Polyglot edge architecture with Rust SIMD vibration pre-filter, Go gateway, Next.js cockpit, and PyTorch AI core deployed across GKE and K3s with SQLite write-ahead caching for 120+ simulated industrial sensors.

Architectural Abstract: Deployed edge anomaly detection (PyTorch Autoencoders) and visual defect inspection (<20ms CNN inference) mitigating part escape defect rates by 31% and preventing $420K in scrap waste. Engineered resilient Kubernetes deployments with automated failover between Google GKE and lightweight CNCF k3s runtimes, backed by local SQLite write-ahead caching for full offline resilience across 120+ simulated industrial sensors.

Key Stack: PyTorch, OpenCV, Rust SIMD, Go Gateway, GKE, K3s, MQTT Broker, SQLite WAL, Jenkins CI/CD.

Runbook: curl -fsSL https://raw.githubusercontent.com/KishorKumarParoi/smart-manufacturing/main/bootstrap.sh | bash

Nexus-AI Multi-Cloud Lakehouse Master Architecture
πŸ” Architecture Blueprint
Code Blueprint Docs


Kishor Kumar Paroi
Multi-Cloud IaC • Distributed Lakehouse • 5-Agent Supervisor 12K req/s • 82% Faster Analytics

Autonomous 5-agent mesh in LangGraph paired with Go 1.22 API gateway (12,000 req/s), DuckDB analytical engine (<120ms queries), and unified multi-cloud Terraform IaC across AWS, GCP, and Azure.

Architectural Abstract: Orchestrated a 5-agent mesh (Data NL2SQL, Business Intelligence, MLOps Supervisor, OWASP Firewall, Voice Alerts) over LangGraph, accelerating turnaround on complex ad-hoc analytics by 82%. Constructed a 3-tier Medallion Lakehouse with DuckDB and Parquet storage achieving sub-120ms dashboard queries. Paired a Go 1.22 API gateway sustaining 12,000 requests per second with full multi-cloud infrastructure codified via Terraform across AWS (EKS, S3, Glue), GCP (GKE, BigQuery), and Azure (AKS, Synapse).

Key Stack: Terraform, AWS, Azure, GCP, DuckDB, LangGraph, Go Gateway, Next.js, Docker Compose.

Runbook: curl -fsSL https://raw.githubusercontent.com/KishorKumarParoi/Nexus_AI/main/bootstrap.sh | bash

Advanced AI Chatbot Architecture
πŸ” Architecture Blueprint
Code Blueprint Docs


Kishor Kumar Paroi
Production Agent Orchestration • Stateful Checkpointing Sub-50ms Recovery • Zero State Loss • FastMCP

LangGraph cyclic state machines with PostgreSQL persistence checkpointers, Model Context Protocol (FastMCP) Human-in-the-Loop gating, BotShield WAF honeypots, and Redis semantic intent caching (<12ms).

Architectural Abstract: Built an enterprise conversational agent using LangGraph cyclic state machines with PostgreSQL persistence checkpointers, empowering fault-tolerant multi-turn thread persistence, dynamic rollbacks, and sub-50ms session recovery under heavy concurrency. Standardized isolated tool execution using Model Context Protocol (FastMCP) with strict Human-in-the-Loop (HITL) approval gates on financial actions (>$50 refunds/cancellations), eradicating 100% of unauthorized transaction risk. Pioneered BotShield WAF middleware featuring honeypot traps (/.env, /wp-admin), scanner fingerprint dropping, and sliding-window rate limiting; intercepted 99.4% of adversarial probes. Operationalized two-stage hybrid retrieval with Qdrant (BGE-Large dense vectors) and BM25 sparse keyword matching with real-time SSE token streaming, elevating response relevance to 94.8%.

Key Stack: LangGraph, PostgreSQL Checkpointer, FastMCP, Qdrant, Redis Semantic Cache, BotShield WAF, Streamlit.

Runbook: git clone https://github.com/KishorKumarParoi/Advanced_AI_Chatbot_Project.git && cd Advanced_AI_Chatbot_Project && ./scripts/deploy.sh

Data Agent Architecture Flow
πŸ” Architecture Blueprint
Code Blueprint Docs


Kishor Kumar Paroi
Hierarchical LangGraph Mesh • Self-Reflective SQL Security Judge • ReAct ETL 96.2% Routing Accuracy • 75% Faster BI

LangGraph supervisor router classifying business queries into specialized SQL and ETL analyst subgraphs, with dynamic schema introspection, independent LLM security judge blocking DDL/DML attacks, and self-healing SQL execution.

Architectural Abstract: Architected a LangGraph supervisor router classifying business queries into specialized SQL and ETL analyst subgraphs, garnering 96.2% routing accuracy across multi-domain operational requests. Instituted dynamic PostgreSQL schema introspection with an independent LLM security judge, sanitizing generated SQL and intercepting 100% of destructive DDL/DML injection attacks. Built autonomous ReAct tool-calling agent for file extraction, pandas data transformation, and sandboxed python execution, compressing ad-hoc data pipeline development time by 75%. Integrated self-healing SQL execution with automatic syntax retry & schema reflection, programmatically correcting 89% of initial query execution errors without human intervention.

Key Stack: LangGraph, PostgreSQL, Multi-Agent Supervisor, Sandboxed Python, ReAct ETL, Streamlit, DuckDB.

Runbook: git clone https://github.com/KishorKumarParoi/Data-Agent.git && cd Data-Agent && ./deploy.sh

Azure ETL Architecture Diagram
πŸ” Architecture Blueprint
Code Blueprint Docs


Kishor Kumar Paroi
Azure Data Factory • ADLS Gen2 • Databricks Spark • Synapse CETAS 1.4B Records • 6.2x Faster Queries • 98.4% Accuracy

Enterprise Lakehouse orchestrating batch extraction via Azure Data Factory (ADF v2), Databricks Apache Spark transformation, Synapse CETAS star schema materialization, and sub-4ms ONNX inference serving.

Architectural Abstract: Architected an enterprise Lakehouse orchestrating batch extraction via Azure Data Factory (ADF v2) dynamic lookup and parallel ForEach loops, landing raw security telemetry into ADLS Gen2 Bronze layer. Conducted large-scale ETL in Azure Databricks (Apache Spark) with Azure AD OAuth 2.0 Client Credentials, cross-joining real-time operational metadata from MongoDB Atlas. Materialized analytical serving layer in Azure Synapse Analytics Serverless SQL using Managed Identity scoped credentials and CETAS (Create External Table As Select) for instant Power BI and Fabric reporting. Built data validation engine with schema drift traps, KNN imputation + RobustScaler, and multi-model benchmark (XGBoost, CatBoost, Random Forest delivering 98.4% detection accuracy).

Key Stack: Azure Data Factory, Azure Databricks, Apache Spark, Synapse CETAS, ADLS Gen2 Delta Lake, MongoDB Atlas, ONNX Runtime.

Runbook: git clone https://github.com/KishorKumarParoi/Advanced_ETL_Project.git && cd Advanced_ETL_Project && ./deploy.sh

AWS SageMaker DAG Pipeline Architecture
πŸ” Architecture Blueprint
Code Blueprint Docs


Kishor Kumar Paroi
SageMaker Pipelines (DAG) • Managed Spot Training • Model Registry 70% Cost Reduction • 2,500 req/s • NIST AI RMF Compliant

Automated SageMaker DAG Pipeline with Managed Spot Training (70% compute savings), holdout quality gating (≥85%), Model Registry promotion, DataCaptureConfig live payload drift detection, and KMS envelope encryption.

Architectural Abstract: Designed an automated SageMaker DAG Pipeline (ProcessingStep → TrainingStep → ConditionStep → ModelStep) with AWS Managed Spot Training, curtailing model training compute costs by 70%. Instituted programmatic quality evaluation gating (≥ 85% holdout accuracy) for promotion into SageMaker Model Registry, deploying to real-time and serverless endpoints with standard inference contracts (model_fn, input_fn, predict_fn, output_fn) and FastAPI fallback. Configured DataCaptureConfig streaming 100% of live inference payloads to Amazon S3 for baseline statistical drift detection, anomaly alerting, and programmatic retraining triggers. Devised multi-model endpoint deployments with autoscaling policies (target tracking at 70% CPU/GPU), sustaining 2,500 req/s with zero cold-start timeouts under spiky production loads.

Key Stack: AWS SageMaker Pipelines, Managed Spot Instances, Model Registry, DataCapture, KMS CMK, S3, Docker, FastAPI.

Runbook: git clone https://github.com/KishorKumarParoi/AWS-SageMaker.git && cd AWS-SageMaker && ./scripts/deploy.sh

Realtime MLOps KServe Architecture Flow
πŸ” Architecture Blueprint
Code Blueprint Docs


Kishor Kumar Paroi
KServe v1beta1 • AWS ECS Fargate • Kubernetes • Terraform IaC • Envoy Sidecars 65% Idle Spend Slashed • Scale-to-Zero

Serverless model inference platform on Kubernetes utilizing KServe v1beta1 scale-to-zero autoscaling, containerized ECS Fargate microservices, Trivy security gating, and Envoy canary routing.

Architectural Abstract: Deployed serverless model inference using KServe v1beta1 InferenceService on Kubernetes backed by Amazon S3 model storage; facilitated dynamic scale-to-zero autoscaling, curtailing idle GPU/CPU compute spend by 65%. Architected containerized microservices orchestration on AWS ECS on Fargate with Application Load Balancers and modular Terraform IaC; accelerated infrastructure provisioning by 80% with zero server maintenance overhead. Hardened end-to-end container security with Trivy CVE scanning and AWS KMS envelope encryption, ensuring 100% programmatic CI/CD gating with zero critical vulnerability deployments.

Key Stack: KServe v1beta1, Kubernetes (EKS), AWS ECS Fargate, Terraform, Envoy Sidecars, Trivy, Prometheus, S3.

Runbook: git clone https://github.com/KishorKumarParoi/realtime-mlops.git && cd realtime-mlops && ./deploy.sh

Rust Git Engine Internals Architecture
πŸ” Architecture Blueprint
Code Blueprint Systems


Kishor Kumar Paroi
Bare-Metal Systems Programming in Rust • Concurrency • Process Pipelines Within 3% of C Git • 73% Lower Peak Heap

Crafted Git internals in pure Rust (SHA-1 object graph, streaming zlib decompression with SIMD vectorization, delta packfiles) paired with async POSIX shell featuring signal multiplexing and I/O redirection.

Architectural Abstract: Matched C Git performance within 3% while guaranteeing 100% compile-time memory safety. Built a custom POSIX terminal shell with asynchronous signal multiplexing, fork/exec process execution pipelines, and file descriptor I/O redirection; clocked sub-microsecond execution latency under heavy concurrent process loads with 73% lower peak heap allocation.

Key Stack: Rust SIMD Intrinsics, POSIX System Calls, Fork/Exec Pipelines, Memory Safety, Async Signal Multiplexing.

Runbook: git clone https://github.com/KishorKumarParoi/codecrafters-git-rust.git && cd codecrafters-git-rust && cargo run

3D Re-Invent Space Station WebGL Platform
πŸ” Architecture Blueprint
Live 3D Platform Code Blueprint


Kishor Kumar Paroi
Three.js • React Three Fiber • GSAP 3.13 • Rolldown Vite 60 FPS WebGL • Sub-120ms Choreography

Interactive 3D WebGL space station and cyber cockpit platform engineered with React 19, Three.js, React Three Fiber, GSAP camera choreography, and Tailwind CSS v4.

Architectural Abstract: Built custom WebGL shaders, floating island GLTF model loader with Draco compression, responsive camera perspective switches, and GSAP timeline animations delivering smooth 60 FPS across desktop and mobile browsers.

Key Stack: React 19, Three.js, React Three Fiber, GSAP 3.13, Vite, Tailwind CSS v4.

Runbook: git clone https://github.com/KishorKumarParoi/re-invent.git && cd re-invent && npm install && npm run dev

Professional Career & Experience
phi0.ai • Forward Deployed AI Engineer July 2026 – Oct 2026 • Europe (Remote)
  • Deployed 5 concurrent autonomous AI agents running cyclic state machines (LangGraph) with dynamic routing, attaining 99.4% task completion under enterprise workloads.
  • Devised a high-throughput Hybrid RAG engine combining dense HNSW vector search, sparse BM25, and multi-stage re-ranking (FlashRank ONNX + Cohere v3.5) with semantic caching (<35ms latency), curtailing hallucinations by 46% and reducing token costs by $18K/month.
  • Codified continuous MLOps flywheels with NeMo Guardrails, automated PII scrubbing, and canary rollouts on Kubernetes, supporting $1.4M in enterprise contracts.
Samsung AI Core / Care Health • Data Engineering & AI Systems Intern March 2026 – June 2026 • Gurgaon / Remote
  • Formulated batch and streaming pipelines via AWS S3, EMR, Athena, PySpark; cut operational overhead by 34% through event-driven Lambda/SQS asynchronous dispatch.
  • Accelerated 8B/70B LLM inference on NVIDIA H100 GPUs using FlashAttention-2 and fused CUDA kernels, realizing 3.8x throughput and reducing p99 latency by 54%.
  • Spearheaded multi-node fine-tuning with PyTorch FSDP and DeepSpeed ZeRO-3 with HIPAA-compliant de-identification, trimming training time by 42% ($14K saved/run).
  • Deployed real-time health insurance claim fraud detection over AWS Kinesis and OpenTelemetry, safeguarding $380K annually across audits.
kepi.ai • Senior Full Stack Engineer Jan 2022 – Sep 2025 (3 yrs 9 mos) • Remote
  • Architected zero-downtime microservices migration to Kubernetes on AWS EKS; scaled platform throughput by 300% while reducing p95 latency by 48% at 99.99% availability.
  • Fortified zero-trust mesh infrastructure with Istio mTLS and fine-grained RBAC, diminishing cross-zone egress expenses by 38%.
  • Designed asynchronous event-driven core using Redis Streams and Celery, sustaining 15M+ daily background jobs with zero message loss.
foodpanda • Full Stack Engineer 2021 – 2022 • Dhaka / Remote
  • Crafted high-traffic checkout microservices with Next.js/Node; trimmed latency by 34% and sustained 100K+ daily peak orders at 99.98% uptime, preventing $120K in checkout cart abandonment.
Open-Source Code & Systems Software

A selection of systems, tools, and repositories authored by Kishor:

  • KoomAI Platform: Sub-120ms WebRTC video communication platform with Electron desktop client, Manifest V3 Chrome extension, and Next.js 16 App Router dashboard.
  • Rust SIMD & Go LLMOps Mesh: High-throughput vector search accelerator utilizing CPU AVX-512 and ARM NEON intrinsics (<0.5ms scoring latency, 50x speedup).
  • Enterprise AI Security & FastMCP: Production agent mesh with BotShield WAF, NVIDIA NeMo rails, Portkey semantic caching, and Microsoft PyRIT adversarial validation.
  • Zomato Lakehouse Pipeline: Medallion streaming lakehouse over Snowflake, PySpark Delta Lake, dbt Core SCD-2, and Debezium CDC feature streaming.
  • Smart Manufacturing Industrial Edge AI: Multi-cloud MLOps with automated failover between Google GKE and lightweight CNCF k3s runtimes, with offline SQLite write-ahead caching.
  • Nexus-AI Multi-Cloud Mesh: Complete multi-cloud IaC with Terraform managing EKS, GKE, AKS, and a 5-agent supervisor coordinating complex ad-hoc analytics.
AI Engineering Learning Journey • 38-Module Curriculum

An exhaustive, production-tested curriculum codified across the AI-Engineering repository. Spanning foundational prompt engineering to agentic cyclic state machines, synthetic data generation, multimodal OCR, and distributed cloud inference:

Phase I: Foundations & Retrieval

  • 00_Setting_Up_Git • Git hygiene, secrets scanning, pre-commit hooks
  • 01_Prompt_Engineering • Few-shot, chain-of-thought, structured JSON outputs
  • 02_Embeddings_and_RAG • Vector similarity, chunking strategies, dense embeddings
  • 03_End-to-End_RAG • Complete naive RAG pipelines with LangChain & OpenAI
  • 04_Production_RAG • Production observability, error handling, token latency
  • 09_Finetuning_Embeddings • Matryoshka representations, domain contrastive loss
  • 13_Advanced_Retrieval • Multi-query expansion, reciprocal rank fusion (RRF)
  • 20_Data_Preprocessing_Embeddings • High-throughput tokenization & chunk cleaning

Phase II: Agents & State Machines

  • 05_First_Agent_with_LangGraph • State graphs, tool binding, conditional edges
  • 06_Multi_Agent_with_LangGraph • Supervisor patterns, agent handoffs, subgraphs
  • 24_LangGraph • Cyclic graphs, PostgreSQL memory checkpointers, rollbacks
  • 25_Multi_Agents • Hierarchical team meshes with specialized tools
  • 26_Agentic_AI • Self-reflective reasoning loops, autonomous planning
  • 27_Production_Agents • Resilient retries, timeouts, deterministic execution
  • 14_Open_DeepResearch • Recursive web crawling, multi-source synthesis
  • 17_On_Prem_Agent • Local air-gapped agent deployment with Ollama & vLLM

Phase III: Evaluation, Fine-Tuning & Vector Databases

  • 07_Synthetic_Data_Generation • LangSmith datasets, LLM-as-a-judge filtering
  • 08_Evaluating_RAG_With_Ragas • Faithfulness, Answer Relevance, Context Precision
  • 10_Finetuning_Reasoning_Model • QLoRA, DeepSeek-R1 style CoT reinforcement
  • 21_Rag_Qdrant_Pipeline • Dense HNSW indexing, payload filtering, scalar quantization
  • 22_Rag_Qdrant_Hybrid_Search • Dense + Sparse BM25 vector hybrid retrieval
  • 23_Rag_Hybrid_Re_Rank • FlashRank local ONNX + Cohere Rerank v3.5 fallback
  • 18_Inference_Quantization • AWQ, GPTQ, GGUF 4-bit/8-bit memory optimizations
  • 19_Testing • Automated unit testing, pytest mocks, evaluation benchmarks

Phase IV: Production Systems & Cloud MLOps

  • 15_Open_Source_Endpoints • vLLM, TGI, self-hosted open-weight LLMs
  • 16_LLMOps • Prompt registry, latency tracking, semantic caching
  • 28_LLM_Engineering • Production engineering best practices, token budgets
  • 30_Forward_Deployment_Engineering • Client-facing integration, production SLAs
  • 32_Agentic_RAG_Project • Full-stack enterprise search with memory & tools
  • 34_Enterprise_Rag_Development • Enterprise RBAC, audit logging, PII redaction
  • 35_Multimodal_Intelligence_System • Vision-language models, document understanding
  • 36_Nemotron_Parse_Mistral_OCR • NVIDIA Nemotron parsing + Mistral OCR extraction
  • 37_Astro_Airflow • Orchestrated batch ingestion DAGs with Astro CLI
  • 38_AWS_Sagemaker • SageMaker DAG Pipelines, Managed Spot, Model Registry

Phase V: Platform Engineering, FastMCP & HITL Governance

  • FastMCP_Tool_Standardization • Decoupled server-side tool execution with typed contracts
  • HITL_Break_Glass_Gates • Human authorization on sensitive operations (>$50 actions)
  • Zero_Trust_BotShield_WAF • Honeypot traps, sliding-window rate limits & IP reputation
  • Semantic_Cache_Gating • Sub-12ms intent caching with Redis & Cosine similarity
  • Multi_Cloud_Terraform_IaC • Single-command deploy.sh and $0.00 clear.sh teardown
  • ArgoCD_GitOps_Delivery • Multi-cluster canary rollouts across AWS EKS & GCP GKE
  • NIST_AI_RMF_Compliance • ISO 27001, SOC 2, HIPAA PII scrubbers & audit trails
  • Distributed_Observability • Prometheus, Grafana, OpenTelemetry, Logfire & LangSmith

Enterprise Flagship Systems & Specialized Repositories (Non-Numeric Modules)

  • AI_Sequrity_Production_App • BotShield WAF, NVIDIA NeMo Guardrails, Portkey, Microsoft PyRIT RedTeaming
  • Advanced_AI_Chatbot_Project • LangGraph cyclic state machine with PostgreSQL checkpoint persistence & HITL
  • Data-Agent • Hierarchical multi-agent SQL router, self-reflective security judge & ReAct ETL
  • Nexus-AI • Multi-cloud Medallion Lakehouse across AWS, GCP, and Azure with 5-agent supervisor mesh
  • Advanced_ETL_Project • Azure Data Factory, ADLS Gen2, Databricks Spark, and Synapse CETAS analytics
  • Advanced_Production_Rag • Two-stage dense HNSW + sparse BM25 + local ONNX FlashRank re-ranking cascade
  • AI_Infrastructure • Production Terraform IaC, EKS/GKE cluster topology, and ArgoCD GitOps pipelines
  • JobHunterAI • Autonomous agentic job scraper, ATS resume tailor & automated candidate briefing
  • RayCodes_AIJobSearch • Autopilot daemon, email drafter, and company reconnaissance crawler
  • Tripmate-Multi-Agent • Collaborative multi-agent itinerary planner with memory checkpointers
  • ai-job-search • Personal career intelligence engine with Trojan Horse tailored outreach proposals
  • dossier • Executive technical dossier and background briefing generator for Principal drills
  • zomato-analysis • 35M+ orders Medallion Lakehouse with Debezium CDC and Feast Feature Store
Technical Skills & Core Pillars
  • Languages & Core Frameworks: Python, C/C++, Rust (SIMD, Actix-web), Go (Gin/Fiber), SQL, Bash, TypeScript, JavaScript, Java (Spring Boot), FastAPI, Express.js, NestJS, Next.js 16, Streamlit.
  • AI, ML & LLMOps: PyTorch, vLLM (Multi-LoRA Dynamic Multiplexing), LangGraph (Cyclic State Machines), LangChain, HuggingFace, LangSmith, LangFuse, DeepEval, TRL, PEFT (RSLoRA, QLoRA), NeFTune, MLflow, Feast Feature Store, Kubeflow v2, AWS SageMaker, KServe, DVC.
  • Big Data & Medallion Lakehouses: Snowflake, Databricks, Azure Data Factory (ADF), Azure Synapse Analytics (CETAS), ADLS Gen2, dbt Core (SCD-2), PySpark, Delta Lake, Apache Kafka, Apache Airflow, Debezium CDC, DuckDB, ChromaDB, Qdrant.
  • Cloud Platforms & Infrastructure (IaC): AWS (EKS, S3, KMS, EMR, Athena, ECS Fargate, Bedrock, SageMaker), GCP (GKE, BigQuery, GCE, VertexAI, Gemini AI), Azure (Databricks, Synapse, Fabric, vLLM, AKS), Terraform, Ansible, Helm, ArgoCD GitOps, Sonatype Nexus, Docker, Kubernetes (GKE, EKS, AKS, K3s).
  • DevSecOps & Observability: NVIDIA NeMo Guardrails, PyRIT RedTeaming (Microsoft), BotShield WAF, SonarQube, Trivy CVE Scanning, Jenkins, GitLab CI/CD, GitHub Actions, Prometheus, Grafana, EFK Stack, OpenTelemetry, Redis.
  • WebRTC & Systems Engineering: WebRTC (SVC, <120ms latency), ElectronJS, Manifest V3 Chrome Extensions, Next.js 16 App Router, Stripe Billing, PostgreSQL, MongoDB Atlas.
Education & Specialized Courses
Rajshahi University of Engineering & Technology (RUET) 2018 – 2023 • Rajshahi, Bangladesh

B.Sc. in Computer Science & Engineering • CGPA: 3.25 / 4.00 (Degree with Honors)
Merit Rank: 483rd Nationwide in Bangladesh

Core CS & Engineering Foundations (RUET): Advanced Data Structures & Algorithms, Object-Oriented Programming (C++/Java), Operating Systems & System Programming, Database Systems (DBMS), Distributed Systems & Cloud Computing, Computer Networks & Internet Protocols, Theory of Computation & Automata, Compiler Design, Software Engineering & System Analysis, Computer Architecture, Artificial Intelligence, Microprocessors & Interfacing, Discrete Mathematics, and Network Security & Cryptography.
Saint Joseph Higher Secondary School 2015 – 2017 • Dhaka Board

Higher Secondary Certificate (HSC) • GPA: 5.00 / 5.00
Merit Rank: 85th in Dhaka Board • National Merit Scholarship

S.J Govt Model High School 2013 – 2015 • Gopalganj, Dhaka Board

Secondary School Certificate (SSC) • GPA: 5.00 / 5.00
Merit Rank: 90th in Dhaka Board • National Merit Scholarship

Certifications, Specializations & System Design Courses

Deep Learning Specialization

DeepLearning.AI & Stanford • Andrew Ng

Neural Networks, Backpropagation calculus, Hyperparameter Tuning, Regularization (Dropout, BatchNorm), Convolutional Neural Networks (CNNs), Sequence Models, Attention Mechanisms, and Transformers.

Machine Learning Specialization

Stanford Online & DeepLearning.AI • Andrew Ng

Supervised Learning (Linear/Logistic Regression, Decision Trees), Advanced Learning Algorithms, Neural Networks, Unsupervised Learning (K-Means, PCA, Anomaly Detection), Recommender Systems, and Reinforcement Learning.

Enterprise Production MLOps & AI Systems

Vivian Aranha, Mayank Agarwal & Krish Naik

End-to-End MLOps Pipelines (CampusX / Krish Naik), Enterprise AI Systems Architecture (Vivian Aranha), Agentic-AI & n8n (Mayank Agarwal), Computer Vision (Sourangshu Paul), AI Data Analyst (Ansh Lamba), Automation & FastMCP (Ed Donner).

Distributed Systems, Big Data & AI Design

Arpit Bhayani, Zach Wilson, Hussein Nasser & ByteByteGo

High-Throughput Distributed Systems & System Design (Arpit Bhayani, Gaurav Sen, Hussein Nasser, ByteByteGo), Big Data Engineering & Medallion Lakehouses (Mayank Agarwal, Zach Wilson), LLM Foundations (Karpathy, Aurimas), and 2000+ LeetCode/Codeforces algorithmic mastery.

Interactive 3D Engineering Platform (re:invent)

3d.kishor.tech • Live Platform • GitHub: re-invent

Architected high-performance 3D portfolio ecosystem with React 19.2, Rolldown Vite, and Tailwind CSS v4; choreographed GSAP 3.13 + ScrollTrigger 3D camera choreography, interactive Bento showcase grids, Supabase PostgreSQL real-time visitor counter with anti-tamper rate limiting, and sub-100ms Vercel Edge CDN delivery.

Achievements, Hackathons & Mentorship
  • Competitive Programming & Algorithms: Ranked 232nd in Codeforces Div-2 (Blue Coder), 5β˜… on CodeChef, 17th in IUPC, Qualified in Meta Hacker Cup Round 3. Solved 2000+ DSA and system design problems.
  • Champion | Technobeast AI Agents Hackathon: Pioneered Qdrant vector optimization with negative embeddings, trajectory matching with transformers, and multimodal API divergence detection.
  • 7th Rank | AI Security Hackathon: Ranked in top tier for real-time prompt injection firewall (97.2% accuracy, <30ms latency).
  • In-Person AI Agents Building Masterclass: Conducted immersive hands-on in-person workshops on architecting autonomous AI Agents, cyclic state machines (LangGraph), tool calling, and production deployment.
  • Technical Mentorship & Community Impact: Mentored 100+ students and engineers across Machine Learning, AI Agents, DevOps, and MLOps, leading code reviews and production architecture design sessions.