// data analyst  →  data engineer  →  agentic ai

I build systems that
show their evidence.

Data analyst and engineer moving into agentic AI, from CCAR regulatory reconciliation at Barclays to a natural-language-to-SQL engine and a full-stack recommendation app. I ship things that are tested, not just demoed.

See Madli, the flagship build → Browse all projects
5
Shipped projects
3+
Years, data & regulatory
scroll
Origin

Why I'm building, not just applying

CS undergrad in India, then Zemoso Technologies, a startup where the competition was real and shipping under pressure wasn't optional. A Master's at UNT came next, not as a pivot but to put formal rigor under instincts I'd already earned on the job.

"I don't just want to exist in this AI race. I want to excel in it."

I was mid-program when generative AI stopped being a research topic and became the industry, and that's when I committed to the data field for real. I took the Data Analyst role at Barclays and went straight into CCAR reconciliation, a regulated environment with zero tolerance for a sloppy pipeline.

Everything since has run in parallel: five shipped projects, each one deliberately closing a gap in the stack agentic AI engineering demands, plus one that's genuinely mine. That's Madli, built and owned end-to-end because a portfolio of tutorials doesn't prove anything a real product does. The Build Theory Fellowship is where I learned to validate before I build, not just prototype in a vacuum.

Grounded in
B.S. Computer Science (India) → M.S. Computer Science, UNT
Day job
Data Analyst, Barclays · CCAR & regulatory reconciliation
How I work
Energetic and competitive, the kind of teammate who takes the lead and finishes what the group started
Right now
Shipping Madli end-to-end as a Build Theory Fellow, toward Data Analyst / Data Engineering / Agentic AI roles
Flagship project
In active development

Madli

A locally-ranked food & travel discovery app that replaces endless scrolling with three honest picks and one reason why. Designed and built end-to-end, from product design through database architecture, security, full-stack implementation, and QA.

View repository ↗
PICK 01
Ranked by real comparisons
Every visit is slotted into a personal list by pairwise comparison, not a star average. New place vs. one you've already ranked, and it finds its spot.
→ hover to see why
PICK 02
Evidence, not just an answer
Every pick shows its sample size and ranking gap, including an honest "not enough data yet" state instead of a fabricated recommendation.
→ hover to see why
PICK 03
Access enforced at the DB, not the UI
Row-Level Security on every table, tested against unauthenticated, wrong-user, and wrong-role access, not just written and assumed correct.
→ hover to see why
This is literally how Madli recommends: three picks, one reason each, evidence attached.
Architecture

Frontend

  • React 19 · TypeScript (strict) · Vite 5
  • React Router v7 · TanStack Query v5
  • Custom design-token CSS system, no Tailwind or UI kit
  • Vitest + React Testing Library

Backend & data

  • Supabase: Postgres 17, Auth, Row-Level Security
  • 17-table schema, 12+ reproducible migrations
  • SECURITY DEFINER functions for audited operations
  • One deployed Edge Function (Deno) for share previews

Design system

  • 3 roles × 2 breakpoints
  • 28 reusable components, documented prop contracts
  • Every UI state audited: loading, empty, error, disabled

Testing & QA

  • Playwright E2E covering happy paths and failure paths
  • RLS tested against wrong-user / wrong-role access
  • 4-phase SDLC with a "verified, not just written" bar
What three roles actually see
G

Guest

Capped, session-only browsing that converts to sign-up, a tuned funnel rather than an afterthought paywall.

U

User

Pairwise ranking, multi-stop trip plans reordered by real shortest-route distance, personalized picks.

A

Admin

Ranking overrides, access-logged sensitive-data review, moderation queues, fine-grained permission grants, enforced by the database rather than a toggle.

What the category keeps getting wrong

Infinite lists, no verdict

  • Yelp/Google/TripAdvisor hand you hundreds of ranked-by-nobody results and call it a decision
  • Madli forces exactly three picks per list, each slotted in by pairwise comparison against a place you've actually ranked before, not a star average

Feeds tuned for engagement, not trust

  • Instagram/TikTok-style discovery optimizes for what keeps you scrolling, not what you'd actually go do
  • Every Madli pick shows its sample size and ranking gap, including an honest "not enough data yet" state instead of a confident guess

Reviews nobody can verify

  • Public review scores mix first-timers, bots, and people who never went back, with no way to tell which
  • Recommendations are grounded in a user's own ranking history, with Row-Level Security and audited admin actions instead of an unmoderated free-for-all

Two distinct uses of AI in this build

As a product feature: recommendations run on a content-based scoring engine today, with a designed LLM-assisted re-ranking layer (Anthropic Claude API) next, deliberately grounded in real ranking history so it can't invent a reason for a suggestion. As the build method: the whole app was built by directing an AI coding agent through a structured four-phase SDLC with specification-grade prompts and a hard "verified before complete" standard.

Also shipped

Four more, all tested end-to-end

The first two were cloned and rerun clean-room, wiped and rebuilt exactly per their own README, before a single claim below got written down. The last two were built fresh for this portfolio, with the same bar: every stat below comes from an actual passing test run.

Agentic AI · Safety

Agentic NL-to-SQL Analytics Engine

Turns a business question into SQL against a live schema. The real engineering problem isn't generation, it's trust. A whitelist validator sits between the LLM's SQL and execution: only a single, schema-valid SELECT is ever allowed to run.

215
Tests passing
0
Unsafe writes executed
PythonFastAPIClaude APIRAG schema grounding
View on GitHub
Regulatory · Automation

Regulatory Reporting Automation Tool

Models the CCAR / FR Y-14Q reconciliation workflow from Barclays: three synthetic source systems, a materiality-threshold reconciliation, and root-cause labeling from variance direction, validated against 14 deliberately planted breaks.

60
Tests passing
14/14
Planted breaks found
PythonSQLSnowflakeStreamlit
View on GitHub
Agentic AI · RAG

Hybrid-Search RAG Engine with Reranking

A two-stage retrieval pipeline: TF-IDF + BM25 hybrid search narrows a corpus to candidates, then a transparent lexical reranker (term coverage, phrase overlap, IDF-weighted rarity) reorders them before an answer is generated. Own test suite included.

27
Tests passing
2-stage
Retrieve → rerank
Pythonscikit-learnBM25Claude API
Publishing to GitHub
Agentic AI · Data Analysis

AI Data Analysis Agent

Chat with a CSV/Excel dataset in plain English. An AST-based whitelist validator, using the same safety pattern as the NL-to-SQL engine, checks every generated pandas snippet before a sandboxed executor runs it on a copy of the data, blocking imports, file I/O, and sandbox-escape chains.

37
Tests passing
13
Adversarial safety tests
PythonPandasAST validationClaude API
Publishing to GitHub
Track record

Experience

Sep 2024 - Present

Barclays

Data Analyst · Whippany, NJ
  • Automated recurring data-validation checks in Python and SQL to cross-verify source-to-target integrity, cutting manual QA time on regulatory deliverables by 70%.
  • Built and maintained ETL pipelines consolidating financial data into centralized Snowflake reporting tables.
  • Reconciled CCAR regulatory datasets with joins, CTEs, and window functions, resolving 25-50 discrepancies per quarterly cycle.
  • Authored data-lineage documentation and reconciliation control frameworks supporting audit readiness for CCAR and FR Y-14Q submissions.
Jan 2023 - Jul 2023

University of North Texas

Graduate Teaching Assistant, Big Data
  • Co-taught a graduate Big Data course covering distributed processing (Hadoop, Spark) and SQL optimization.
  • Led lab sessions for 40+ graduate students on SQL and Python assignments.
Dec 2021 - Nov 2022

Zemoso Technologies

SQL Developer · Hyderabad, Telangana
  • Wrote and optimized SQL queries, stored procedures, and views, improving query performance by 40%+.
  • Performed end-to-end ETL for downstream analytics systems and translated requirements into database schemas.
Jan 2023 - May 2024

M.S., Computer Science

University of North Texas, Denton TX
  • Database Systems · Big Data & Data Science · Data Analytics · Algorithms · Artificial Intelligence
Toolbox

Skills

Data Engineering

SQLPython / PandasPySparkETL designSnowflakePostgreSQLOracle

Analytics & BI

Data reconciliationData validationPower BI (DAX)TableauStreamlit

Governance & Regulatory

CCARFR Y-14QBCBS 239Data lineageAudit readiness

AI & Full-Stack

LLM / RAGFastAPIReact / TypeScriptSupabaseRow-Level Security
Credential: Case Study 002

Build Theory Fellowship

Day 0: “I want to become an entrepreneur. I don’t know where to start.” Day 24: Madli, built, deployed, and being tested by real users. This is the fellowship that turned a technical skill set into a shipped product, officially documented as Build Theory’s Case Study 002.

Build Theory Fellowship ID card: Madhupriya Allam, Fellow 1.0, Fellow ID BTF-008

Fellow 006 · ID BTF-008

She came in with the technical foundation: a CSE undergrad, a master’s, self-built projects, but no product idea and no structured way to evaluate one. What the fellowship installed wasn’t coding ability. It was the discipline to validate a problem before committing to build it.

DAY 0 → idea validation → 20+ interviews → PRD & MVP → design system → AI-directed build → DAY 24 → shipped
Idea validation 20+ customer interviews PRD & MVP scoping AI-native build
24
Days, idea to shipped
20+
Customer interviews (12+ strangers)
23
V1 scope areas defined
7
Artefacts owned, not just a certificate

“People don’t need another list. They need help making a decision.”

That single interview insight is what turned Madli from a generic discovery app into a three-picks-and-a-reason product. The interviews didn’t just validate the idea, they changed its scope: personal-vibe filtering over just occasion, time-aware speed for working professionals, tone and aesthetic for Gen Z segments.

The method, repeatable on the next problem
01
Notice
02
Validate
03
Frame
04
Define
05
Design
06
Build
07
Ship
Credentials

Certifications

Certified ScrumMaster (CSM)
AWS Certified Cloud Practitioner
AWS Certified AI Practitioner (AI/ML)
Let's talk

Open to Data Analyst, Data Engineering & AI/Agentic roles.

Based in Jersey City, NJ. Reach out directly. I reply fast.

madhupriyaallam0802@gmail.com (469) 512-3052 github.com/madhupriya08 Download résumé