Aiden Mak

ReviewLensAI — Case Study

↓ Case Study

ReviewLensAI

Reading lens

Info

Stop reading 200 reviews. Ask them a question — and click straight through to the review that proves the answer.

Role

Personal Product

Creator / AI Engineer

Mar 2026

Stack

React

Supabase

Pinecone

RAG

Proof

181/181 Vitest tests and 6/6 promptfoo LLM evals passing

4 ingestion modes — CSV, paste, URL, screenshot — into one searchable evidence base

Live on Vercel: ask a question, click the [Review N] badge, see the source

Live Demo ↗GitHub ↗

Hundreds of reviews, no time to read them — and generic AI summaries hallucinate.

01 — Core Workflow05 Films

Film — ingest_csv.mp4

01 — Ingest From Almost Anything

Drop a CSV in any column layout and the extractor maps the columns itself. Every path ends in the same editable preview table, where bad rows get deleted before they pollute the data — then one click commits them, and every review becomes searchable evidence for the chat.

CSV in any layout — columns mapped automatically

Film — ingest_paste.mp4

02 — Paste the Messy Stuff

Paste a blob of text where one review says "*****", the next says "2/5" and a third uses star emoji — the extractor normalizes all of them into the same preview table, and pasted URLs auto-detect their platform. Messy pasted text stops being "data we ignore."

Mixed rating formats normalized in one paste

Film — ingest_image.mp4

03 — Screenshots Become Data

Upload an app-store screenshot and vision extraction reads reviewer names, star glyphs, dates and Verified badges off the pixels — into the same editable preview table as every other path. The reviews that only existed as images join the searchable, chattable evidence base.

GPT-4o Vision reads reviews off a screenshot

Film — summary.mp4

04 — Instant Analytics, No Run Button

Ingestion computes the analytics on the spot — rating distribution, sentiment breakdown with percentages, date range, per-product dashboard stats. No configuration step, no "run analysis" button: by the time the reviews are in, the shape of opinion is already on screen.

Distribution, sentiment, date range — automatic

Film — reviews_table.mp4

05 — Find and Inspect Any Review

Full-text search, one-tap star filters, pagination — and any row click opens the Evidence Drawer: full text, visual stars, verified badge, source badge, helpful count. It's the same drawer the AI's citations open, which is the point — human browsing and AI evidence share one UI.

Search, filter, open the Evidence Drawer

02 — The AI Layer03 Films

Film — citations.mp4

06 — Answers With Receipts

Questions are embedded and matched against this product's reviews only; GPT-4o answers from the retrieved set and nothing else, streaming token by token. Every factual claim carries a [Review N] badge — click it and the full source review slides in. Out-of-scope questions get a fixed refusal.

Claim → citation badge → source review

Film — chat_skills.mp4

07 — Seven Analyst Lenses

Skill pills switch the analysis mode — Features, UI Bugs, Sentiment, SWOT, Pricing, Executive Summary — each injecting a specialised directive with its own conversation history. The Pricing lens isolates cost complaints naming individual reviewers; the Executive lens compresses 200 reviews into three C-suite sentences.

Pricing lens, then the Executive summary

Film — insight.mp4

08 — A Board-Ready Report in One Click

The Insight tab runs three AI workers in sequence — one finds up to 6 distinct themes, the next distils up to 8 friction points, the last turns both into up to 10 prioritised actions with rationales — and renders collapsible sections with HIGH/MED/LOW badges, PDF download, and a clipboard checklist.

Themes → FAQs → prioritised actions → PDF

Full product walkthrough — narrated (6 min)

03 — Business ValueWhat Changes

Trust

Citations turn AI output from an opinion into an index over evidence — the claim you can't click through to a source never ships.

Time

200 reviews become a themed report with prioritised actions in under a minute, instead of an afternoon of skimming and three remembered quotes.

Coverage

Screenshots and messy pasted text stop being "data we ignore" — every review joins the same searchable, chattable evidence base.

No usage metrics are claimed — the product is pre-revenue and the demo data is synthetic.

Engineering Deep DiveArchitecture · Decisions · Lessons

Want to see how it's built?

The product story ends here — what follows is the engineering. One rule produced most of it: every demo clip was recorded by a scripted browser driving the real product, so each recording doubled as an end-to-end test. The camera earned its keep.

For the technically curious: 5 Supabase Edge Functions, GPT-4o + text-embedding-3-small, top-8 cosine retrieval over per-product Pinecone namespaces, 181/181 unit tests across 24 files, 6/6 promptfoo LLM evals — every number verified against the running system (2026-07-24), and the fixes the camera forced.

01 — AI ImplementationEngineering Lens

Structure over Prompts

The scope guard has two layers. Layer 1 is structural: every product's vectors live in their own Pinecone namespace, and the namespace is resolved server-side from Postgres — the client never supplies it. Even a hostile client cannot retrieve another product's reviews. Layer 2 is instructional: the system prompt requires grounding in retrieved reviews only, mandates [Review N] citations for every factual statement, and pins a fixed refusal for out-of-scope questions. The guard survives prompt injection because the layer that matters isn't a prompt.

Text First, Evidence Attached

chat-rag streams answer tokens over SSE, then emits a separate citations_ready event carrying full metadata for the retrieved reviews — text renders immediately, and badges become clickable the moment evidence arrives. Retrieval is top-8 by cosine, temperature 0.3, last 10 turns of history; the prompt states how many reviews were retrieved out of how many total, so the model can explain its own retrieval when asked.

Three Workers over One Mega-Prompt

The insight pipeline is three sequential GPT-4o workers with tight JSON output schemas, each consuming the previous worker's output — themes feed FAQ extraction, both feed action planning. Small prompts with verifiable structure instead of one mega-prompt: when a section is wrong, you know which worker to fix.

Vision into the Same Schema

GPT-4o Vision reads review screenshots into the same structured schema as CSV extraction — including counting filled versus empty star glyphs and mapping Verified badges, instructions added after the camera caught it miscounting. One schema for four ingestion paths means one preview table, one validation layer, one test surface.

02 — Engineering & ArchitectureSystem Map

The request path is deliberately boring: four ingestion routes — CSV, paste, URL, screenshot — converge on the same extraction schema and land in Supabase Postgres; embed-reviews writes each product's vectors to its own Pinecone namespace; a question hits chat-rag, embeds, retrieves top-8 by cosine, and GPT-4o streams the answer with the evidence trailing it. generate-insight chains the three workers into a report you can hand to a board.

Chat-RAG

Features

UI Bugs

Sentiment

SWOT

Pricingisolates cost complaints

Executive Summarythree C-suite sentences

Skill fan-out — one chat, seven analyst directives (six pictured), each with its own history

Ingestion

extract-reviews — GPT-4o fn-callingextract-image — GPT-4o Vision

Data

Supabase PostgresPinecone — namespace per product

Chat

chat-rag SSETop-8 cosine retrievalcitations_ready

Insight

ThemerFAQ BuilderAction Planner

Frontend

React 19 + ViteTailwind 4SSE reader + jsPDF

Proof

181/181 Vitest6/6 promptfoo evalsPlaywright recorder as E2E

System map — five Edge Functions between four ingestion routes and one evidence-grounded chat

03 — Journey to Completion04 Stories

The Live Demo Was Quietly Dead

Step zero — curl the deployment — found the Vercel frontend serving HTTP 200 while its Supabase project returned NXDOMAIN from every DNS resolver: the free-tier project had been reclaimed after inactivity. The old README even predicted it. Every clip was therefore recorded against the local Supabase stack with real GPT-4o and Pinecone calls, and the restore runbook — new project, db push, secrets, five function deploys, ~30 minutes, zero code changes — is written up in the handoff.

Test Suite Rescue

13 of 173 tests were failing before recording could start, none for product reasons: Node 26's experimental localStorage global shadowed jsdom's; three tests expected a placeholder from an older UI iteration; five mocked an API path the component had moved off — one passing by accident. All fixed → 173/173, and the regression tests added for the on-camera bugs below grew the suite to its final 181/181.

Combined Citations Were Dead Text

First hero take: the answer's first paragraph ended in [Review 1, Review 4] rendered as inert plain text while single citations around it were clickable badges — the badge regex only matched [Review N]. Fix: parse citation groups into one badge per review, regression tests for both the full and shorthand forms, re-recorded with badges clickable on camera. The same pass fixed GPT-4o's markdown bold rendering as literal asterisks.

Vision Miscounted the Stars

The synthetic screenshot's two-star reviewer came back as rating 3, and two Verified badges were dropped. Fix: explicit glyph-counting and verified-badge instructions in the vision prompt plus schema passthrough; Postgres now shows rating 2 and correct flags, re-recorded. A cousin bug — the summary proudly displaying "Ingestion Method: Unknown" because the label map had no image key — was a one-line fix with a parameterised test over all four ingestion methods.

04 — What I Learned04 Lessons

A Dead Demo Fails Silently

A 200-serving frontend told everyone the product was live while every data call failed. Monitoring the backend of a free-tier demo — or a scheduled ping — is part of shipping a portfolio.

The Camera Is the Cheapest E2E Harness

Four real bugs surfaced only when a scripted browser drove the real product — three of them invisible to 181 green unit tests.

Model Output Is an Interface with Drift

Combined citations and markdown bold are the LLM equivalent of an API changing its response shape — render layers need to be liberal in what they accept.

Structural Guarantees Beat Prompt Promises

Server-side namespace resolution makes cross-product leakage impossible rather than discouraged.

05 — Roadmap & CollaborationFuture Work · Contact

Restore the hosted backend (runbook ready) · auth + RLS (design doc exists) · PDF ingestion with page-level citations · URL pagination · sentiment trends over time · competitor benchmarking · Slack/Jira push · model routing and multi-provider support. All labeled future work.

Open to Collaboration

AI/full-stack engineer, Toronto — open to AI engineering roles and selected agency projects. If your product decisions sit on a pile of unread reviews, I'd love to compare notes.

Email ↗LinkedIn ↗Live Demo ↗

Reading lensLive Demo ↗