Agent online · e-commerce & Shopify

Profit forensics,
on autopilot.

Now hunting: RTO spikes

LossLens is an AI agent that finds why your store lost money. Drop in your exports: a deterministic engine computes the numbers, an AI analyst reasons through 8 steps, and a verifier catches anything it can't trace. Out comes the exact root cause and a ₹ fix plan.

Launch the agent →Replay a sample case

Designed & engineered by S Surya Narayan

agent://losslens · demo replay
$
◆ Campaign × Gender◆ Campaign × Product saturation◆ Campaign × Geo◆ Catalogue contraction◆ Carrier failure × RTO◆ Cost-price COGS spike◆ Workhorse defunding◆ Geo-exclusion bypass◆ ROAS floor removed◆ OOS-forced migration◆ Silent RTO◆ Spend bloat◆ Campaign × Gender◆ Campaign × Product saturation◆ Campaign × Geo◆ Catalogue contraction◆ Carrier failure × RTO◆ Cost-price COGS spike◆ Workhorse defunding◆ Geo-exclusion bypass◆ ROAS floor removed◆ OOS-forced migration◆ Silent RTO◆ Spend bloat

// architecture

The AI reasons. Code does the math.

Most AI analysts make up numbers. LossLens splits the work: a deterministic engine owns every core figure, the model owns judgement, and a verifier holds both to account.

01

Intake

files → sources

Drop CSV exports or paste tables. Each source is recognised and mapped to the steps it unlocks. Gaps are declared up front.

02

Engine

code · deterministic

Parses the PnL, runs the benchmark flowchart, and computes deltas, flags and ₹ impact. Same input, same numbers, every run.

03

Reasoning

AI · 8 steps

Works through changelog, fulfilment, audience, product, geo and campaigns, ruling each in or out with evidence.

04

Verifier

guardrails

Re-checks every core figure against the engine, replaces anything untraceable, and flags weak conclusions.

05

Report

dashboard + JSON

A cross-cut root cause, a prioritised action plan with owners, the full case study, and a machine-readable sidecar.

inside the reasoning layer

S1

Benchmark selection

Searching ±10/20/30% spend bands across 6-month, 12-month and full-history windows

S2

Problem metric

Comparing RTO, S/GMV, COGS and Cancel against thresholds and sizing ₹ impact

S3

Changelog correlation

Matching price, cost, visibility and OOS changes to the flagged metric

S4

Fulfilment checks

O2S, S2A, FAD, NDR and silent-RTO signals

S5

Audience

Gender, age and platform shifts, with mechanical chunks

S6

Product mix

Concentration, anchor SKUs and the demand-saturation fingerprint

S7

State / geo

Geo expansion vs spikes in previously clean states

S8

Campaign cross-cut

Workhorse defunding, targeting compliance, and campaign × dimension

// guardrails

Built to not hallucinate.

Each failure mode seen in real loss cases has a matching check. Live checks run in code on every report. Framework rules are enforced inside the reasoning method.

● live check01

GMV back-calculated from S/GMV

Every GMV figure must match a PnL cell. Mismatches are replaced.

● live check02

Numbers the AI made up

All four core metrics are recomputed in code and swapped in if they differ.

● live check03

Loss week used as a benchmark

The flowchart only accepts >5% profit or 3–5% break-even, inside the spend band.

● live check04

Impact assigned to an improved metric

₹ impact is stripped from any metric that didn't worsen.

● live check05

Wrong headline problem

The primary lens is checked against the largest flagged ₹ impact.

● live check06

Single-factor 'root cause'

A root cause must cross two dimensions, or it gets downgraded.

○ framework rule07

Campaign RTO used as the account flag

The method compares account-level PnL RTO only. Campaign RTO is evidence.

○ framework rule08

Carrier failure mistaken for geo

Spikes in previously clean states trigger an FAD/NDR carrier review.

○ framework rule09

Demand saturation missed

Per-SKU spend growth vs order growth and CPO ratio are checked.

// simulations

Four stores. Four different failures.

Each fictional store is built so the obvious guess is wrong. Load one and the agent's intake fills in for you.

// launch

Brief the agent.

Four questions for context, then drop your exports. The agent shows which steps your data unlocks before it starts.

Leave blank and the engine picks the most recent loss week.

A hypothesis. The agent will confirm it or overturn it.

Pick all that apply.

0 / 500,000
⇪

Drop CSV exports here or click to browse

PnL · campaigns · changelog · audience · state · product · fulfilment · up to 1–2 months (daily rows OK)

step unlock matrix

PnL required

0
1
2
3
4
5
6
7
8
· Weekly PnL· Catalogue changelog· Fulfilment / carrier· Gender / age / platform· Meta weekly trend· Product PnL· State breakdown· Payment / coupon· Campaign PnL
> awaiting data · load a simulation or drop files

// new · section studio

Fixed the leak? Rebuild the storefront.

Describe a Shopify section in plain English: hero, product grid, testimonials, FAQ, trust badges. The AI writes production-ready Liquid with theme-editor settings. Preview it live, refine it by chat, and download.

Open Section Studio →

{% schema %}
{ "name": "Hero banner",
"settings": [ … 12 settings ],
"blocks": [ button × 2 ],
"presets": [ … ] }
{% endschema %}
✓ schema valid · ✓ presets · ✓ scoped CSS

// roadmap

Where this goes next.

● Live · v1

Web agent

Guided intake, deterministic engine, 8-step reasoning, verifier, dashboard and JSON sidecar.

● Live · beta

Section Studio

AI-generated Shopify sections with live preview, chat refinement and Liquid download.

○ Next

Batch mode

10 stores per run with a priority queue ranked by budget and loss depth.

○ Next

Direct connectors

Pull from Shopify, Meta Ads and courier APIs, with no manual exports.

○ Later

Auto-routing

Catalogue and logistics findings become tickets for the team that owns the fix.

○ Later

Golden-set evals

Every rule change must hold agreement on a library of solved cases before release.

// faq

Before you trust it.

Only the weekly seller PnL is required: spend, GMV, RTO%, Spend/GMV%, COGS%, Cancel% and Net Profit% per week. Every extra block (changelog, campaign PnL, gender/age/platform, state, product, fulfilment, Meta weekly trend) unlocks another step. Anything missing is flagged as a data gap, never invented.

Three levels deep. Level 1 is the broken metric (for example, RTO is up 17pp). Level 2 is a single factor (for example, more orders going to men). Level 3 crosses two factors to name the actual mechanism: for example, campaign × gender, where new 'Women Only' campaigns delivered 64% of orders to men and caused doorstep refusals. Level 3 is where you can act.

Not the core numbers. A deterministic engine written in code parses your PnL, picks the benchmark, and computes every metric delta, flag and ₹ impact. The AI receives those figures as fixed inputs. A verifier then compares the AI's output against the engine and swaps in the computed value wherever they differ. You can see each check, and the engine trace, above the dashboard.

With a strict flowchart. The tool looks for profitable (>5%) weeks first, then break-even (3–5%) weeks. It tries spend bands of ±10%, then ±20%, then ±30%, and looks back 6 months, then 12 months, then the full history. It prefers 4 consecutive weeks over 1. Loss weeks never qualify. If nothing fits, the comparison is labelled DIRECTIONAL.

Only metrics that got worse are counted. Each percentage point of change is multiplied by a fixed share of loss-week GMV: RTO 0.8%, S/GMV 2.0%, COGS 1.0%, Cancel 0.5%. These numbers rank where the money went. They are not an exact accounting reconciliation.

This app has no database. Your paste is sent once to a free AI model (Pollinations, or Google Gemini if configured) to produce the analysis and then discarded by the app. Don't paste data you're not allowed to share.

Yes. That's why the framework forces every number to come from your data, marks estimates, and lists the data that would change the conclusion. Treat the output like a sharp analyst's first draft, and verify the top action before you act on it.