Open to work · Bengaluru · Open to relocation

VickyYadav

Senior AI EngineerAgentic AI ArchitectAuthor

I build AI systems that listen, see and act — multi-agent assistants inside the Ola app, live voice pipelines that dub launches across languages, a health assistant at Zyvelor AI, and the data platforms that train them. Eight years in production, from fintech and IoT to Google Ads.

softmax(QKᵀ/√d)·Vhover a row to read it
0years shipping production systems
0Ola scooter centers on CCAI
0processed · 33% lower compute
0ATMs secured by VTAS
0Realtor.com pages indexed

01 · Selected workAgents, voice, vision & data

Ola · Krutrim AI · 2024–2025

Krutrim AI Chat — an agentic assistant inside the Ola app

Led the assistant end to end. People shop, order food, book cabs, file complaints and get recommendations by talking to it. ReAct agents on LangGraph plan each turn, call tools against Ola services, keep state across steps and fall back cleanly when a tool fails.

  • MCP servers with Composio as the integration layer, so agents reach third-party apps and internal services the same way.
  • Benchmarked CrewAI, Agno and PocketFlow against LangGraph and chose per use case on latency, cost and reliability.
  • Java Spring Boot, React/Next.js, ScyllaDB and PostgreSQL; live analytics in New Relic, MoEngage and an ElasticSearch dashboard.
LangGraphLangChainReActMCPComposioSpring Boot
One agent turn
user message→planner · ReAct→tool router→MCP · Composio→Ola services→state + reply
Live dubbing pipeline
stream in→ASR→translate→TTS→face match→MediaLive
Zyvelor AI · 2025 — now

Kaira AI — a health assistant, built from scratch

People log meals by typing or snapping a photo, and ask questions about their body-composition (BCA) and continuous glucose monitoring (CGM) data. I designed the LLM pipeline: query understanding, food-database retrieval, nutrition estimation and device-data context.

Multimodal LLMRAGVision-languageHealth data
Google · 2025

Ads Solutions Architect

Search Ads 360 specialist. Architected backend integrations for campaign and ads management at scale for enterprise clients including Binghatti and Warta, and designed measurement and automation across Google’s ads products.

SA360Ads APIsMeasurementAutomation
Ola · Krutrim AI

Real-time AI dubbing for live launches

ASR, translation and TTS as FastAPI microservices on AWS Elemental MediaLive, so launches streamed in several Indian languages at once. YOLO face recognition tracked who was speaking and synced the translated voice to the right face.

ASRTTSNMTYOLOFastAPI
Ola · Krutrim AI

Contact Center AI

An agentic playbook platform behind WhatsApp presale chatbots for Ola scooter centers — booking test drives and closing purchases, shipped through Argo CD with New Relic monitoring.

4,000+centers
Agent playbooksWhatsAppArgo CD
Ola · Krutrim AI

Training data at 22 TiB

A master–worker Spark pipeline over 22 TiB that cut compute cost by a third, and a re-architected OBELICS pipeline extracting 50–100 million Indic text–image pairs from 230M+ sources for multimodal training.

−33%compute cost
Apache SparkOBELICSMultimodal data
Ola · Krutrim AI

LoRA fine-tuning for Indic tasks

Fine-tuned open-source LLMs with LoRA (PEFT) in PyTorch and Hugging Face Transformers for domain and Indic-language tasks — better answers at a fraction of full fine-tuning cost.

LoRAPEFTPyTorchHugging Face
QSS Technosoft · 2019–2020

VTAS ATM surveillance

Designed and led an 8–10 person team on the IoT backbone now securing ATMs across India — Kafka, RabbitMQ, Apache Ignite, Spark Streaming and Node.js sockets talking to DVRs, cameras and intrusion panels.

50,000+ATMs
KafkaRabbitMQSpark StreamingIoT
Brillio · 2021–2023

Search & price intelligence

SEO engineering for Realtor.com with Spark sitemap ingestion indexing 2M+ pages on Google and Bing; Dell’s competitor price pipeline with AI table extraction from web, PDF and Excel; a Bing Maps data-center locator for Microsoft.

SparkNext.jsAI extractionPlaywright

02 · Research in progressTwo papers on the desk

Working titles · drafts in 2026
In progress

Context rot in long-running AI agents

Agents that run for hours fill their context with stale tool output, superseded plans and old turns — and quietly get worse. This paper studies how that decay shows up, and which memory, compaction and retrieval strategies hold it off.

Question: when does more context start to hurt an agent?

In progress

A faster database for AI agents

Agent workloads read and write in a pattern of their own — short memories, tool results, embeddings and state checkpoints, all on the hot path of every step. This work designs a storage layer around that pattern to cut the latency an agent waits on.

Question: what does a database look like when its main user is an agent?

The Secret Language of Machines — front cover
03 · The book · First edition, 2026

The Secret Language of Machines

How numbers learn to read, reason, paint and sing. A story of the mathematics behind modern AI, written for everyone who ever stared at an equation and decided it was not meant for them — every symbol pronounced out loud.

It climbs from a single scalar through vectors, probability and calculus to attention, transformers, diffusion, voice models, scaling, post-training and agents — and ends with a GPT built from scratch.

29 chapters220 pagesPart V · build agents with Agno & PocketFlow

04 · CareerEight years, newest first

  1. Nov 2025 — now

    Senior Data Scientist · Zyvelor AI

    Kaira AI health assistant: multimodal food logging, BCA and CGM questions, the full LLM pipeline.

  2. Jul — Nov 2025

    Ads Solutions Architect (Product) · Google

    Backend integrations and Search Ads 360 automation for enterprise clients including Binghatti and Warta.

  3. Jan 2024 — Jul 2025

    Senior Full Stack Lead · Ola (Krutrim AI)

    Krutrim AI Chat, ReAct agents, MCP and Composio, CCAI, live dubbing, LoRA fine-tuning, 22 TiB data pipelines.

  4. Feb 2023 — Jan 2024

    Senior Full Stack Engineer (Product) · Zluri

    The Overview module in React, faster Node.js APIs, a documentation summarizer and a process-flow assistant.

  5. Dec 2021 — Feb 2023

    Senior Full Stack Engineer · Brillio

    Realtor.com SEO platform, Dell price intelligence, Microsoft CILMA analytics.

  6. Dec 2020 — Dec 2021

    Full Stack Lead · RattanIndia Finance

    Led 12 developers and 2 QA engineers to build the loan portal from scratch, with Paynimo payments and Digio e-KYC.

  7. Jan 2019 — Dec 2020

    Software Development Engineer · QSS Technosoft

    VTAS ATM surveillance, the GoUpendo backend for a US startup that raised $5M, and the Uptake Medical dashboard.

05 · StackWhat I reach for

Real rigid-body physics · grab, throw, stack
highlightsAI & agents voice & visionbackend datacloud & frontend

drag a block · throw it · shake the box

    06 · RecognitionAlong the way

    • Author, The Secret Language of MachinesFirst edition, 2026 · books.threadlane.in
    • Founder, Threadlane AI LabsApplied generative-AI products · ai.threadlane.in
    • Winner, Pesto Award for Best Project LeadPesto Tech
    • Best Debut EmployeeQSS Technosoft, 2019
    • Mentor to 20+ engineersFull-stack engineering at Pesto Tech

    07 · EducationFoundations

    • Master of Computer ApplicationsCDAC (GGSIPU), New Delhi · 2016–2019
    • Bachelor of Computer ApplicationsMagadh University, Patna · 2013–2016

    Building agents, voice or vision systems?

    I’m open to work. Based in Bengaluru, open to relocation. The fastest way to reach me is email.

    vy4693@gmail.com
    +91 93153 89865