ALL SYSTEMS OPERATIONAL Jugal Sheth / Mission Control
Resume
Senior AI Data Engineer — New York City

I build production AI systems. Not demos.

Welcome to the Data Factory — the portfolio as a place. Every building is a system I shipped; every amber beacon is a failure that taught me a principle. Fly around, or take the list view.

53,400+
lines, solo platform build
77
API routes in one system
$0
variance vs ERP, 8 subsidiaries
280/280
reports recovered
◉ PRODUCTION DASHBOARD THE 5-METRICS FRAMEWORK
p95 latency412ms
Cost per request$0.018
Task success rate94.2%
Fallback rate3.1%
User correction rate2.4%
procurement-agentOPERATIONAL
cdc-monitorOPERATIONAL
hr-sync-agentOPERATIONAL
0 incidents this week · the 5-metrics framework
Zone 01 · Start here

Entrance Plaza

This is a portfolio you walk through. The big hall on your left is the flagship build. The glowing river is a live CDC pipeline. The amber beacons are incidents — click one. Use the map above, arrow keys, or take the tour.

drag to orbit · scroll to zoom · ← → switch zones · click a building or beacon
Senior AI Data Engineer — New York City

I build production AI systems. Not demos.

I turn fragile data pipelines and AI agents into systems that survive real users, real costs, and silent failures — with parity checks, guardrails, and proof, not vibes.

Snowflake · CortexAWSPython · TypeScriptdbtCursor · Claude Code
◉ PRODUCTION DASHBOARD THE 5-METRICS FRAMEWORK
p95 latency412ms
Cost per request$0.018
Task success rate94.2%
Fallback rate3.1%
User correction rate2.4%
AGENT FLEET
procurement-agent · approvalsOPERATIONAL
cdc-monitor · change data captureOPERATIONAL
hr-sync-agent · nightly syncOPERATIONAL
0 incidents this weekframework from reel #09 · 5 metrics
0+
lines of code — solo build of a production platform
0
API routes shipped in one system
$0
variance vs the ERP across 8 subsidiaries at close
0/0
expense reports recovered after a silent join failure
● LIVE FROM GITHUB 45 public repos 1 follower · counts verified Nov 9, 2026
01 // Systems & the principles they taught me

Systems & the principles
they taught me.

Not a task list. Each system left behind a principle I now apply everywhere — the story is just the proof.

Flagship · Solo build

Principle — If it can't be reconciled to the ledger, it isn't done.

A procurement platform built end-to-end: approvals, vendor workflows, and ERP sync designed so every dollar could be traced. I owned architecture, data model, and delivery — no handoffs to hide behind.

Next.jsAWS AmplifySnowflakeERP integration
53,400+ lines · 77 routes · 7+ integrations
ConstraintFinance needed $0 variance at close, not “close enough.”
DecisionBuilt reconciliation into the pipeline — parity checks on every sync, not a month-end audit.
Outcome$0 variance vs the ERP across 8 subsidiaries.
CDC Pipeline

Principle — Identity is the contract between systems.

A webhook CDC pipeline where the same job had to mean the same thing in two systems of record. When IDs drift, everything downstream lies quietly.

ConstraintTwo systems, one job, no shared key you could trust.
DecisionCross-system job identity with idempotent replay.
OutcomeDuplicate execution eliminated; replays became safe.
HR Migration

Principle — Migrate the parity check first.

An HR system moved to a cloud warehouse — 17 nightly jobs rebuilt, every domain validated before cutover, not after complaints.

Constraint16 API domains, zero tolerance for missing people.
DecisionParity-verified migration, domain by domain.
Outcome16/16 domains matched. 17/17 jobs green — and actually correct.
Headcount Model

Principle — “How many people did we have on March 3rd?” is a hard question. Model for it.

A point-in-time headcount model over 2.8M rows. The first version answered “today” perfectly and history wrong. A window-function rewrite made any date answerable — because headcount has two dates: when it happened, and when you learned about it.

2.8M rows
point-in-time correct
2-date model
effective + knowledge date
02 // Principles, learned the hard way

Green jobs lie.

Every incident below passed its checks. The dashboards were green. The data was wrong. These are the stories I tell so your team doesn't have to live them.

Principle 01

Green jobs lie: a job can succeed while the data is wrong.

Silent

An approval queue looked healthy — but an exact-match join across two systems was silently hiding work. People were waiting on approvals that never appeared.

BEFORE
85%
AFTER
100%
The fixLEFT JOIN + fuzzy matching instead of exact-match. Visibility went 85% → 100%.
Principle 02

A round number is a fingerprint.

Pagination

Salaries were stuck at exactly 5,600. Not 5,612. Not 5,589. Exactly 5,600 — which is 28 batches × 200. When a count lands on a round number, suspect pagination, not luck.

STUCK
5,600
SIGNAL
28×200
The lessonMonitor field population, not job success. A full page that returns exactly the page size is a warning, not a result.
Principle 03

Claim before you send.

Duplicate

952 emails went to 242 people. One admin got 8 copies. Two workers reading the same queue will both send — unless one claims the work atomically first.

SENT
952
PEOPLE
242
The fixAtomic claim-before-send in the outbox. Idempotency isn't a feature — it's the design.
Principle 04

Never trust an exact-match join across two systems of record.

$181M+ flagged

An INNER JOIN quietly ate 280 expense reports — and $181M+ got falsely flagged in the process. The query didn't fail. It just returned less truth.

LOST
280
BACK
280/280
The lessonMonitor field population, not job success. Count what should be there, not just what arrived.
03 // Break my agent

Try to break it.
Controls are the product.

Most demos show what an agent can do. This one shows what it refuses to do — and why that refusal is the feature.

EVAL PLAYGROUND · GUARDRAIL TEST
Adversarial eval — live in your browser
● AGENT ONLINE
PRESET ATTACKS

Every run checks: instruction hierarchy → policy → human-approval gate → audit log. If any layer fails, the action never executes.

Agent: procurement-agentSTATUS: OPERATIONAL
// Reasoning trace will appear here. // Pick an attack on the left to run an eval. ✓ Guardrails loaded: instruction hierarchy, approval gate, audit logging, least-privilege tools ! Human review required for any transaction approval
VERDICT
Awaiting eval
No attack run yet.
EVAL SCORE
—
guardrail hold
04 // Ask Jugal — RAG over the corpus

Ask the work,
not the résumé.

Answers are matched from real writing and career docs. If there's no evidence, it says so — it doesn't invent.

SUGGESTED QUESTIONS
HOW THIS WORKS

Keyword-matched retrieval over a small curated corpus for this concept. Citations show the source type. No model call, no hallucinated biography.

Hey — I'm the corpus version of Jugal. Ask me what I build, how I catch silent failures, or what “production-ready” actually means. [system]
05 // 5-metrics self-assessment

Is your agent
production-ready?

Slide to your real numbers. The verdict names the first thing to fix — because “almost ready” is how incidents happen.

50mstarget < 800ms3000ms
$0.001target < $0.05$0.20
40%target ≥ 90%100%
0%target < 10%40%
0%target < 5%30%
YOUR VERDICT
94/100
Production-ready ✓

All five metrics are inside production thresholds. Keep monitoring — readiness is a habit, not a launch.

SHAREABLE RESULT
My agent scored 94/100 on Jugal's 5-metrics check — production-ready.
06 // Stack

The stack I ship on.

Snowflake

Warehouse, parity checks, point-in-time models at scale.

Cortex

AI inside the warehouse — where the data already lives.

AWS

Amplify, Glue, Lambda — boring infra, reliable outcomes.

Python

Pipelines, evals, and the glue between every system.

TypeScript

Full-stack product surfaces people actually use.

dbt

Tested transformations. If it isn't tested, it isn't trusted.

Cursor

AI-assisted shipping — with review, not blind acceptance.

Claude Code

Agentic coding workflows for production codebases.

CURRENTLY EXPLORING → Agent eval harnesses · MCP tool design · Cost-aware routing · Human-in-the-loop UX
07 // Authority — building in public

Proof compounds
in public.

REELS

@jugal.builds

Production AI, no demo theatre. The 5-metrics framework and agent-failure breakdowns started here.

WRITING

LinkedIn, daily

Tech + AI + economics, in plain language. One real system lesson per post.

SPEAKING

Next: the stage

Proposals submitted — none confirmed yet: NY dbt Meetup, Thu Dec 10, 5:30-8pm EST, Sigma Computing, 1 Madison Ave — "The Test Your dbt Project Is Missing" (submitted Nov 9, awaiting reply); NYC School of Data — "Green Jobs Lie — and Public Data Lies the Same Way" (submitted Nov 9, awaiting reply); ContainerDays & AI Context NY (submitted Nov 9 via Sessionize, CFP closes Dec 4, awaiting reply).

JUDGING

Give back

Hackathons and AI awards judging — the goal is to be in the room where good work gets recognized.

08 // Contact

Building in public.
Speaking, writing, collaborating.

No form that goes nowhere. Pick a channel — I actually reply there.

CURRENT FOCUS
Production AI systems that earn trust — then keep it.

Senior AI Data Engineer in New York City. I write and speak about the unglamorous work that makes AI reliable: reconciliation, evals, guardrails, and catching the failures nobody alerts on.

Connect on LinkedIn →