Restaurant intelligence · B2B SaaS
Disha Analytics
Compresses two analyst-days of restaurant research into a five-minute, LLM-generated intelligence report.
The problem
Independent restaurant owners make operational decisions without structured data. They know something is wrong — reviews mixed, margins thin — but cannot afford consultants and have no tool that turns scattered signal into a clear action plan.
How it works
- An analyst enters a restaurant name and city. Disha scrapes 8 public data sources — including TripAdvisor, Yelp, and Google Places — and runs 8 specialist AI agents: competitor, financial, review, menu, hypothesis verifier, validation, synthesis, and owner narrative.
- It produces a ranked action plan, a competitor benchmark, a hypothesis checklist, and a financial-signal section.
- A belief graph structures the evidence chain — Claims map to Signals via weighted Edges, so every conclusion is traceable back to a specific raw observation rather than a black-box LLM output.
- Three delivery modes: Public Scan (public data only), Validated Report (owner-supplied financials), and Owner Walkthrough (a decision deck).
- A three-tier ingestion architecture — format adapters → validators → canonical models — keeps the pipeline robust across messy sources.
- Built end-to-end: Django 5 REST API, React 18 + Vite frontend, a Celery + Redis async pipeline, and PDF reports via WeasyPrint, deployed on Hetzner.
Architecture
Restaurant name + city
optional: owner financials, menu, POS export
8 public data sources scraped
Reviews
Menus
Competitors
Maps
Pricing
Financials
6 specialist AI agents
Decision-ready report
Ranked action plan
Competitor benchmark
Hypothesis checklist
Three delivery modes
Public Scan
Validated Report
Owner Walkthrough
Name + city → public sources → specialist agents → synthesis → a decision-ready report.
The journey
A consulting engagement compresses months of scattered data into a week-long analysis sprint. Disha runs the same process in five minutes — and then keeps going.
The analyst types a restaurant name and city — that's the entire input
The pipeline figures out the rest. It scrapes 8 public sources in parallel, benchmarks against local competitors, surfaces the highest-demand dishes, and identifies the financial signals buried inside public review patterns.
The intelligence brief surfaces — ranked, specific, with named prices
Not a dashboard full of charts to interpret. A brief. With a first action: raise the Chicken Shalimar price from €14.95 to €17.00. A margin estimate. A scenario calculation. The kind of output a senior analyst would produce after a week of work.
The brief — specific action, tagged by effort and impact, with financial scenario
The belief graph traces every conclusion back to raw evidence
Customer Voice, Competitive, Menu, Financial — four signal columns, each node linked to the raw observation that generated it. Every claim is citable. There is no black box. The restaurant owner — or a sceptical stakeholder — can follow the chain from conclusion back to source.
Belief graph — 19 nodes, 9 hypotheses, every claim traceable
Labs auto-generates experiments from the report findings
The system doesn't stop at analysis. It designs what to test next — 7 experiments generated automatically from the report hypotheses, each with a measurement approach, a duration, and success criteria.
Labs — 7 auto-generated experiments, each with a measurement framework
Any hypothesis can be designed into a rigorous experiment in plain English
Type what you want to test. Disha writes the protocol — switchback design, 4-day duration, 63 observations per arm, exact stop conditions. A statistically valid experiment designed from a single sentence.
Plain English input → AI designs the protocol
Switchback design, 4 days, 63 obs/arm — ready to run
Highlights
- 8 specialist report agents + 4 Labs agents
- 8 public data sources (TripAdvisor · Yelp · Google Places)
- Belief graph: Claims / Signals / Edges — every conclusion traceable
- 3 delivery modes
- ~5-minute report
Stack
Why it matters
The consulting instinct from Amazon, productised — take a complex data problem and compress it into a decision-ready output under time pressure.