Deepanshu Kalra — Data & AI Engineer

centerpiece

the agent debugger

Your AI agent doesn't have a model problem. Pick a layer, read the failure mode, then run the diagnosis. From the essay — the full stack, explained.

agent-debugger // v1.0 ● live

layer 01 // prompt

the instruction

does

can't

failure signature

> ingesting incident report

> layer 01 PROMPT instruction well-formed [OK]

> layer 02 CONTEXT 14 chunks retrieved, recency verified [OK]

> layer 03 HARNESS no approval gate on tool `deploy` [FAIL]

> layer 04 LOOP no stop condition after 12 turns [WARN]

> layer 05 PROTOCOL 3 servers connected via MCP [OK]

> ─────────────────────────────

> DIAGNOSIS: not a model problem.

> failure in the scaffolding — layer 03 HARNESS.

> know the layer → know the fix.

by the numbers

telemetry

0

years in data engineering

0

users reached by my data systems

0

product features powered at Edge

0

humans following along on LinkedIn

what i do

system modules

From raw events to decisions — and now to agents. The throughline of my career is making data trustworthy and useful.

mod_01

pipelines & warehousing

Modern ETL, data lakes, and warehouses on GCP, Azure, and AWS — built to be reliable at billion-user scale.

mod_02

analytics platforms & BI

Self-serve dashboards, dimensional modeling, and metrics that leadership actually uses — including SVPs at Google.

mod_03

AI agents

Gemini-powered agents and workflow automation — like the AI workbench my team uses daily, inside Google Chat.

career

service record

Twelve years across the full data stack, from BI consulting to billion-user platforms. Open a row for the full detail.

Working in the Devices & Services group on data strategy and technical implementation. Building modern pipelining, warehousing, and reporting solutions — plus Gemini-powered AI agents — used by teams across the organization, up to SVP level.

awardInnovation Award, Google

Core data engineering team for Microsoft Edge. Built telemetry and customer analytics frameworks on top of Chromium's engine, data lakes on Azure, and a self-serve dashboarding system that enabled analysis for 50+ features. Worked directly with feature PMs and senior leadership.

awardStar Award, 2019

Analyzed US TV viewership from raw television/STB data. Rated programs and channels with complex time-zone normalization and proprietary algorithms — data sold to third parties and used to optimize channel bundling across TiVo operators.

Worked directly with customers including Unilever (UK) and Novartis (Germany) across BFSI, HR, and healthcare. Trained new hires in SQL and data technologies.

Designed a data mart, ETL, and multidimensional solution for an internal Microsoft site.

awardSpot Award — teamwork & the usage cube

writing

field notes

Occasional essays on data engineering, AI agents, and building systems that survive production.

essay_01

Your AI agent doesn't have a model problem

Four layers, four failure modes — why most agent failures live in the scaffolding, not the model. Plus the connectivity stack: MCP, xMCP, and agent-to-agent.

oct 12, 2026 · read

essay_02

Big Data Architectural Patterns: Lambda, Kappa, Zeta

Batch and speed layers, the append-only immutable log, and the seven-component enterprise stack — three patterns for taming volume, velocity, and variety.

from the linkedin archive · read

essay_03

Best places to start learning SQL for free

Twelve free resources — interview prep, interactive tutorials, window functions, and practice platforms. From my most-shared LinkedIn post.