$ whoami  →  agentic_ai_engineer.py

VAIBHAVKATHAIT

Forward Deployed AI Engineer/Agentic AI & Enterprise Automation

I design and ship multi-agent systems, RAG pipelines and LLM infrastructure — from enterprise platforms for Verizon and Western Alliance Bank to open-source agent tooling built in my own time.

📍 Remote · India 3+ yrs in production GenAI GCP GenAI Leader Certified
Scroll
01 / About

Designing the systems that let agents act.

AI Engineer with 3+ years of experience building enterprise AI applications, agentic workflows, data automation and full-stack software for business and financial-services use cases — including work for Western Alliance Bank and Verizon, AI agent hiring workflows for Landing Point, and an SFMC automation platform.

I build across the whole lifecycle, from understanding stakeholder needs and designing workflow logic to API integration, implementation and delivery. Strongest in Python, SQL, FastAPI, LangChain, LangGraph, MCP, RAG and multi-tenant backend design, with cloud exposure across Google Cloud and AWS.

What I care about most is putting practical, cost-aware AI systems into real operational environments — and explaining any of it to a non-technical stakeholder.

0
Years building enterprise AI and data systems
0
Concurrent client environments on one multi-tenant service
0
Languages supported, fully offline (Local Murmur)
02 / Capabilities

capabilities.json

~/vaibhav/skills.jsonread-only
"agentic_ai_&_llm_apps":
LangGraphLangChainOpenAI Agent KitModel Context ProtocolTool / Function CallingAgent Workflow DesignRAGEmbeddingsPrompt & Context EngineeringLLM Evaluation
"software_engineering_&_apis":
PythonC++RustSQLFastAPINode.jsReactREST APIsPostgreSQLSQL ServerMulti-tenant ServicesEnterprise API Integration
"google_cloud_&_data_pipelines":
Cloud SchedulerDataflowBigQueryCloud RunIAMGoogle AI ServicesKubernetes Fundamentals
"aws_&_cloud_delivery":
Amazon BedrockAWS LambdaAmazon S3Amazon SageMakerDockerAzure DevOps CI/CD
"ai_document_&_language_workflows":
Speech / Voice AnalysisNERDocument Data ExtractionIntelligent Document ProcessingUnstructured Data Ingestion
"data_engineering_&_reliability":
ETL PipelinesPySparkMetadata-driven WorkflowsSource ReadinessHoliday-aware SchedulingAuditabilityPipeline Reliability
"machine_learning_&_deep_learning":
PyTorchHugging Face TransformersDiffusersLoRA Fine-TuningStable DiffusionVector EmbeddingsScikit-learn
"developer_&_observability_tools":
CursorClaude CoworkGPT-5.3 CodexLangSmithMLflowWeights & BiasesChromaDB
OPENVISION.DEMO
Built to run unsupervised.

// AgentForge ships with its own guardrails — token-cost limits, iteration caps and dynamic model routing — so an agentic loop knows how to exit safely, not just how to start.

03 / Experience

Where the systems have run.

Agentic AI Engineer
Capital Numbers Infotech Ltd. · Remote · Jan 2026 – Present
Developing an agentic platform that automates Salesforce Marketing Cloud workflows, connecting a custom Model Context Protocol server to the SFMC APIs. Designed a multi-tenant service for 5+ concurrent client environments with asynchronous Pub/Sub processing, decoupling incoming requests from downstream execution while keeping tenants isolated. Translating client requirements into AI-enabled workflows and owning delivery from solution design onward.
Associate Software Developer
Agreeya Solutions · Delhi, India · May 2024 – Oct 2025
Built GenAI tooling and data automation for enterprise clients including Verizon and Western Alliance Bank, covering SRE incident analysis, developer onboarding and AML data operations. Worked across the delivery lifecycle — ingestion, embedding pipelines, retrieval, agent orchestration, testing and deployment — and wrote design docs and led sessions on RAG and LangGraph patterns for team reuse.
Full Stack Developer
PCCS · Delhi, India · Jul 2023 – Dec 2023
Built React and Node.js interfaces and optimised backend REST APIs for internal business applications. Delivered a market-research platform integrating LLM APIs and sentiment analysis with real-time dashboards, used by 10+ internal stakeholders.
04 / Featured Work

Personal builds.

Shipped outside the day job — the projects I build to test ideas the client work doesn't leave room for.

Tiny VLA + FPGAVLA
A small vision-language-action model built from scratch: it looks at a picture and follows a typed instruction like "push the red block left" to steer a robot gripper in simulation. 85.6% success on 500 unseen situations, matching the expert program it learned from, at roughly 8-9x the speed of the usual step-by-step approach. A timing component on an FPGA keeps the moves in step when AI answers arrive late.
VLAPyTorchFPGARobotics
chaos-bringer↗
A chaos monkey for agent frameworks. It fuzzes, fault-injects and red-teams LangGraph, AutoGen, ADK or MCP/A2A agents, judges what the agent actually does, and turns every finding into a replayable regression test. Published on PyPI (pip install chaos-bringer); free by default, running on local Ollama.
PythonPyPIOllamaRed-teamingMCP
Black Box↗
A privacy-first flight recorder for your Mac. It keeps a rolling five-minute buffer of screen, metrics and processes; when something crashes it freezes the evidence and asks a local LLM for a root-cause analysis. Nothing leaves the machine.
Local LLMOllamaRoot-cause Analysis
Sylph↗
Feather-light, fully on-device voice-to-text: press a hotkey in any app, speak, and the words are typed for you. Powered by the 16.9 MB Cactus Whistle speech model inside a Tauri 2 shell, so audio never leaves the machine.
TauriRustPythonOn-device ASR
Back in Time↗
An agent skill that makes AI-agent work replayable and reversible: a plan with an undo for every step, plus a trace of what actually happened, so any run can be rolled back or replayed with new parameters.
Agent SkillsPythonReproducibility
JevFlash↗
An Apple Silicon replication of NanoJev's non-generative decision models: instead of writing tokens, a Qwen3 backbone scores a fixed set of candidate answers in a single forward pass. Ported from CUDA to MPS and CPU.
PyTorchQwen3MPS
mario-wasm↗
Super Mario World 1-1 in 100% Rust, compiled to WebAssembly: physics, tile collisions, enemy behaviour and a chiptune soundtrack synthesised live with Web Audio, playable in the browser at 60 FPS.
RustWebAssemblyWeb Audio
slm-recommendations↗
A small-language-model query–item relevance model, reimplementing DoorDash's search-ads architecture to rank query/item pairs without a full-size LLM in the loop.
PythonSLMSearch Ranking
Local Murmur↗
Fully offline voice dictation for Apple Silicon — whisper.cpp with Metal acceleration paired with a local Llama 3.2 for grammar cleanup. 99 languages, zero cloud dependency, no subscription.
whisper.cppOllamaMetal
AgentForge↗
Open-source, drag-and-drop builder for local AI agent workflows. Topological execution engine, live SSE debugging, per-node model routing, and guardrails on token cost and iteration count.
FastAPIReactOllamaSSE
RUST-AGENT-MEMORY-MVP↗
An MVP for testing how agents behave under different memory-management strategies — a small, fast substrate for poking at recall and decay written in Rust.
RustAgent Memory
Recommendation↗
A semantic movie recommender — Sentence Transformer embeddings indexed with FAISS for nearest-neighbor search, served through a Streamlit front end.
FAISSSentence TransformersStreamlit
Hybrid_db↗
A small database engine built from scratch in Python, exploring what it takes to give JSON documents and SQL tables a shared storage layer.
PythonDatabase Engine
Anime Style LoRAHF
A fine-tuned SDXL LoRA adapter published on Hugging Face, improving anime-style generation quality and facial consistency while preserving the base model's composition.
SDXLLoRADiffusers
05 / Research

Published work.

Zenodo Preprint · Apr 2026
Scaling Spectral Richness: Residual Self-Attention Mechanisms for Genre-Conditioned Audio Diffusion on Small Datasets
Designed and trained an 89.5M-parameter genre-conditioned spectrogram diffusion model from scratch on GTZAN using a single T4 GPU. Introduced zero-initialized residual self-attention and showed it improves harmonic richness even when MSE loss is unchanged — revealing a gap in standard audio-generation evaluation metrics.
DOI: 10.5281/zenodo.19761415 ↗
Kaggle Notebook · Companion Study
Comparative Ablation Study: Residual Self-Attention vs. Baseline in Spectrogram Diffusion
A fully reproducible ablation study comparing models with and without residual self-attention, isolating the architectural contribution to spectral quality — training pipeline open-sourced alongside it.
View notebook ↗
06 / Certifications

Credentials.

Google Cloud Generative AI Leader
Google Cloud
Issued Sep 2026
Verify badge ↗
Lakehouse Fundamentals
Databricks
Certified
// education — B.Tech, Amity University, Noida
07 / Contact

Building an agent
system worth trusting?
Let's talk.

Open to conversations about agentic AI, GenAI infrastructure, and interesting problems in general.