Case study

AI-powered systematic literature review platform.

MettaByte helped turn a complex academic research workflow into a full-stack AI product: from research question to multi-database search, deduplication, screening, overrides, and PRISMA-ready exports.

Plan an AI system See AI automation

Industry

Academic research

Build type

AI web app

Status

MVP scaling

Disclosure

Anonymized

Client name, institution, real review data, and private corpora are intentionally omitted. The case study focuses on product architecture, engineering problems, and safe benchmark figures.

Systematic review pipeline

Search, deduplicate, screen, export

MVP-ready

Databases

Benchmark records

3,600

Dedup recall

100%

AI screening decision

Structured, reviewable, override-safe

Paper abstract

Evaluate the title and abstract against inclusion, exclusion, quality, and research-question criteria.

Decision

Include

Maybe

Exclude

Every decision includes confidence and a plain-English justification for researcher review.

Problem

Systematic reviews take months

Researchers must search multiple databases, deduplicate thousands of records, screen every abstract, and export results for standardized reporting.

Challenge

The hard part was reliability

The work was not simply connecting an LLM. The system needed correct metadata handling, auditability, rate-limit resilience, and researcher control.

Result

A scalable MVP foundation

The platform moved from prototype into MVP preparation, with production-hardening work completed before broader release.

What we built

An end-to-end SLR workflow, not a chatbot.

The platform had to respect how academic reviews actually work: source traceability, researcher judgment, structured exports, and repeatable review history.

Multi-database search

Parallel search across academic databases, with query adaptation and normalized records across different schemas.

AI protocol generation

Drafted Boolean search strings, inclusion and exclusion criteria, and structured review framing from a research question.

Five-pass deduplication

Combined DOI, title/year, author, deposit-DOI, and fuzzy matching to handle messy academic metadata.

LLM screening workflow

Screened titles and abstracts against review criteria with structured INCLUDE, MAYBE, or EXCLUDE decisions.

RIS/BibTeX support

Allowed uploaded citation files for sources that could not be searched directly through an API.

PRISMA-ready exports

Produced review exports and per-source counts needed for standardized literature review reporting.

Technical difficulty

Academic APIs had different query syntax, pagination, field names, rate limits, and abstract availability.

Duplicate detection had to work even when DOI, year, and author metadata disagreed across databases.

LLM screening needed concurrency, timeouts, rate-limit backoff, and rule-based guardrails to avoid silent misclassification.

Researchers needed manual override controls because AI can assist screening, but researcher judgment must remain authoritative.

Authentication had to support a simpler development path while keeping a route open for academic SSO integration.

Architecture

Data model

Search APIs

Security

Next.jsReactTypeScriptTailwind CSSFastAPIMongoDBOpenAIOllama supportsentence-transformersGitHub Actions

Safe results

Public proof points without exposing the research.

These figures describe engineering benchmarks and hardening work. They do not reveal client identity, institution, review topics, paper corpora, or private research data.

Benchmarked deduplication against 3,600 live records from 7 academic databases.

Improved known duplicate-pair recall from 90.3% to 100% after fixing metadata failure modes.

Ran deduplication in roughly 0.10 to 0.19 seconds on a 1,200-record corpus.

Engineered AI screening for 30 concurrent LLM calls with automatic backoff and per-paper timeout tracking.

Resolved 47 production-hardening issues before MVP release preparation.