MANEETSHAH

Maneet Shah · Software engineer

Full-stack products.
Reliable backends.
Applied AI.

Backend Engineer at RevSave.ai. Founder & engineer of Cadence.
Computer Science at UC Irvine · Expected September 2027.

Explore my projects ↘View résumé ↗Email me ↗LinkedIn ↗GitHub ↗

I build across the stack—from authenticated APIs and data pipelines to the interfaces people use.

Backend
Webhooks · API contracts · PostgreSQL
Product
Authentication · Dashboards · Data ingestion
Applied AI
LLM evaluation · Colorimetry · Search
Off the clock? Deal me in. ↓

Selected work.

What I built, what I owned,
and what improved.

Founder & engineer
2025 — Present

Cadence

Athlete readiness, built for coaches.

A full-stack platform that turns athlete check-ins into Train / Modify / Recover guidance.

Built
Authentication, ingestion, PostgreSQL scoring and coach dashboards.
Result
Check-ins under 30 seconds. Row-level security across 25 tables.
Visit the product ↗Replay the decision ↗
Applied AI project

SkinLens

Color analysis with a faster AI pipeline.

Combines deterministic CIELAB colorimetry with structured LLM grading.

Built
A hybrid analysis pipeline and optimized authentication path.
Result
21 → 3 API calls. Analysis under 11 seconds. Hot-path auth: 407 → 13ms.
Replay the decision ↗
Information retrieval

UCI-ICS
Search

Search across tens of thousands of pages.

Built
On-disk inverted index, multi-way merge and term-range partitioning.
Result
Queries under 300ms.
Explore the code ↗

A few strong hands.

Selected GitHub projects. Explore the data, models, and systems behind the work.

View GitHub ↗

Loading the project index…

GitHub snapshot: . Each demo identifies whether it uses published data, a documented example, or a simplified illustration.

Engineering experience.

RevSave.ai

Backend Engineer · August 2026–Present
Webhook integrations, replay protection, and idempotent revenue attribution.

SharkNinja

Data Analytics Intern · 2025
Semantic modeling and a governed NL-to-SQL benchmark with ≥90% accuracy.

AdvaMed

Data Intern · 2024
Expenditure forecasting, from 8–10 days of manual work to same-day results.

ByPeople

QA Intern · 2023
27 defects identified, including six critical issues fixed before release.

A little poker.
A lot of decisions.

Off the clock, I play poker.
Deal through a hand to see how I think.

Pre-flop

Start with what you know.

Before choosing an approach, understand the problem and what is still unknown.

Scroll to reveal · or choose a stageReplay the decisions ↓

Replay the decision.

Three projects. The problem, my contribution, and the result.
Choose a card to look closer.

Cadence

The problem
A coach needs a useful readiness signal without a long athlete check-in.
My contribution
Built authentication, data ingestion, PostgreSQL scoring and coach dashboards.
The result
Check-ins under 30 seconds; row-level security across 25 tables.
Show the tradeoff

A short check-in reduces effort for the athlete, but provides less context than a detailed assessment. Train / Modify / Recover gives the coach a concise signal to interpret.

A question to test next

How well does the readiness signal match coaches’ decisions over time?

Explore the live product ↗

SkinLens

The problem
An analysis pipeline made 21 API calls; hot-path authentication took 407ms.
My contribution
Combined deterministic CIELAB colorimetry with structured LLM grading and optimized the authentication path.
The result
21 → 3 API calls. Analysis under 11 seconds. Hot-path auth: 407 → 13ms.
Show the tradeoff

Deterministic color measurement and model-based grading do different jobs. This separation makes the measurement step explicit, while model outputs still require evaluation. Fewer calls alone do not establish accuracy.

A question to test next

Does grading remain consistent across lighting conditions and repeat runs?

Read the résumé summary ↗

Results documented in my résumé; source code and benchmark data are not linked here.

RevSave.ai

The problem
Webhook retries can cause the same event to be processed more than once.
My contribution
Implemented webhook/API contracts, HMAC verification with a 300-second replay window, and idempotent revenue attribution.
The result
Repeated requests do not trigger duplicate calls or conversions.
Show the tradeoff

Signature verification checks authenticity; idempotency handles duplicates. A replay window limits the age of accepted requests, so delayed delivery also needs explicit handling.

A question to test next

What happens when a valid event arrives after the replay window?

Read the résumé summary ↗

Results documented in my résumé; source code and benchmark data are not linked here.