CostKari

1. One-liner

AI recipe-costing copilot that shows independent Indian restaurants which dishes lose money — before the month ends.

2. Trend signal — why now?

Three things converged in the last 12 months to make this ripe:

Food costs are crushing independent restaurants. The average Indian restaurant runs food costs at 40–45% of revenue when the healthy range is 30–35%. Most don’t even know. A restaurant doing ₹50 lakh annually wastes ₹3–6 lakh on invisible food cost leakage — spoilage, over-portioning, mispriced dishes. That’s not a rounding error; it’s the difference between surviving and shutting down.

AI inference got cheap enough for ₹499/mo unit economics. GPT-4o-mini, Claude Haiku, Gemini Flash — all dropped below $0.10/million input tokens in early 2026. A WhatsApp-based food cost agent that processes daily purchase inputs, runs recipe cost calculations, and sends margin alerts costs under ₹30/month in inference. That was impossible 18 months ago.

WhatsApp Business API matured for transactional workflows. With 96% smartphone penetration in India and the API now supporting rich interactive messages, buttons, and list pickers, you can build a full recipe-costing workflow inside WhatsApp — no app install, no training, no IT department. India’s restaurant owners already live on WhatsApp for supplier orders. Meet them there.

Provenance:

• Signal 1: Indian restaurants run 40–45% food costs vs 30–35% healthy range, ₹3–6L annual waste per restaurant — toyaja.com — 2025
• Signal 2: AI inference costs dropped below $0.10/M tokens (Claude Haiku, GPT-4o-mini, Gemini Flash) enabling sub-₹500/mo SaaS unit economics — industry pricing pages — Q1 2026
• Signal 3: 96% WhatsApp penetration among Indian smartphone users; WhatsApp Business API now supports interactive transactional workflows — waba.nxccontrols.in — 2026
• Signal 4: Existing recipe costing software (MarketMan $149–199/mo, meez, CrunchTime) priced for US chains, not Indian independents — marketman.com — 2026
• Signal 5: 95% of Indian restaurants are single-owner operations; independent operators control 65% of the foodservice market — mordorintelligence.com — 2025 Category: Underserved niche + Geographic arbitrage

3. The opportunity

Recipe costing software exists — MarketMan ($199/mo), meez, CrunchTime, Apicbase — but it’s all built for US/EU chain restaurants with English-speaking managers, desktop workflows, and $500+/mo software budgets. The 300K+ independent restaurants in India (and millions more informal ones) cost their recipes on the back of a notebook, in Excel, or not at all. They literally don’t know which dishes lose money.

The gap: there is no WhatsApp-native, Hindi/regional-language, ₹499/mo recipe costing tool purpose-built for an Indian restaurant owner who does ₹20–80 lakh/year in revenue and has zero IT infrastructure. Petpooja and Posist have basic inventory features, but recipe-level food cost analysis is bolted on, clunky, and still requires desktop access.

CostKari is a focused AI copilot that does one thing: it tells you your real food cost per dish, flags the losers, and alerts you when supplier price changes break your margins. All via WhatsApp. No app. No desktop. No training.

4. Target market

• Primary customer: Owner-operator of an independent restaurant (1–3 outlets) in urban/semi-urban India doing ₹20–80 lakh annual revenue. Cuisines: North Indian, South Indian, biryani houses, Chinese-Indian, multi-cuisine. Typically employs 5–20 staff.
• Why they buy: They know money leaks but can’t see where. They’ve tried Excel and given up. They over-portion, mis-price delivery items, and absorb supplier price hikes without adjusting menu prices. End of month, the bank balance doesn’t match the order count.
• Rough TAM reasoning: 300K+ registered independent restaurants in India. Even the top 10% (30K restaurants willing to pay for software) at ₹499/mo = ₹18Cr ($2.1M) ARR. Expand to cloud kitchens, catering, and dhabas — the addressable base is 100K+.
• Why now for them: Swiggy/Zomato commissions (18–30%) have compressed delivery margins. LPG prices jumped ₹195/cylinder in April 2026. Ingredient inflation is persistent. The owners who survived COVID are now dying of margin erosion. They need visibility into dish-level profitability yesterday.

5. Product sketch (MVP)

• WhatsApp-based daily purchase log: Owner sends a photo of the supplier bill or types “tomato 10kg 400rs” — AI extracts items, quantities, prices
• Recipe builder: One-time setup per dish via guided WhatsApp flow — “What goes into your butter chicken? How many portions does it make?”
• Live food cost dashboard: WhatsApp message with per-dish cost %, color-coded (green/amber/red), updated whenever purchase prices change
• Margin alerts: Automatic WhatsApp notification when a dish’s food cost crosses 35% threshold — “Your paneer tikka food cost jumped from 31% to 38% because paneer went from ₹320 to ₹400/kg”
• Menu pricing suggestions: “To keep paneer tikka at 32% food cost, increase menu price from ₹280 to ₹310 or reduce portion by 15g”
• Hindi + English interface (expandable to Tamil, Telugu, Kannada, Marathi)
• Simple web dashboard for owners who want a visual overview (optional, not required for core value)

6. AI angle — what’s load-bearing

AI is doing real work here, not decoration:

1. Bill parsing: OCR + LLM extracts line items from handwritten/printed supplier bills photographed on WhatsApp. Indian supplier bills are messy — mixed Hindi/English, inconsistent formats, hand-scrawled. This is exactly the kind of unstructured input that LLMs handle well and rule-based systems can’t.
2. Recipe cost calculation with unit conversion: Owner says “half kilo paneer” and “250ml cream” — AI normalizes units, maps to purchase prices, handles yield loss (e.g., 1kg raw chicken = 700g usable), and calculates per-portion cost.
3. Proactive margin monitoring: AI watches purchase price changes across all recipes and surfaces only the actionable alerts — “3 dishes went from green to red this week because onion prices spiked 40%.”
4. Natural language interface: No forms, no dropdowns, no training. The owner talks to CostKari like they’d talk to an accountant. “How much does my biryani actually cost me?” gets a real answer.

Remove the AI and you’re back to a spreadsheet that nobody updates. The AI is the product.

7. Localization angle

This is an India-first play by design. The localization is the moat:

• Language: Hindi-first, English support. Phase 2 adds Tamil, Telugu, Marathi, Kannada — covering 80%+ of restaurant owners. Global recipe costing tools are English-only.
• Payment: UPI Autopay for subscriptions. No credit card required. ₹499/mo hits the sweet spot — less than one day’s food waste savings.
• Distribution: WhatsApp is the operating system of Indian small business. No app store discovery problem, no download friction, no “please open the desktop app.”
• Pricing context: MarketMan is $199/mo (~₹17,000/mo). CostKari at ₹499/mo is 30× cheaper. Even Petpooja at ₹1,000/mo doesn’t offer dedicated recipe costing.
• Supplier bill formats: Indian supplier bills (kaccha bills, handwritten receipts, mixed-language invoices) are a specific OCR challenge that global tools don’t handle.

8. Business model — path to $1M–$5M ARR

• Pricing: ₹499/mo per restaurant (Starter). ₹999/mo adds multi-outlet, waste tracking, and menu engineering recommendations (Pro).
• ACV: ₹6,000–12,000/year ($70–140)
• Rough math to $1M ARR: 7,000 restaurants × ₹1,000/mo average × 12 = ₹8.4Cr (~$1M). Achievable in 18–24 months with strong referral loops.
• Rough math to $5M ARR: 25,000 restaurants × ₹1,400/mo average × 12 = ₹42Cr (~$5M). Requires geographic expansion beyond metros + cloud kitchen segment + pro-tier upsell.
• Expansion path: Add waste tracking (actual vs. theoretical consumption), supplier price benchmarking (“you’re paying 15% more for chicken than restaurants in your area”), and eventually become the procurement layer — aggregate purchasing for better rates.

9. Go-to-market wedge — first 100 customers

1. Restaurant associations and mandi networks: India has active restaurant owner WhatsApp groups in every metro. Join 50 groups in Mumbai, Delhi, Bangalore, Hyderabad. Share a 60-second demo reel showing “your butter chicken actually costs you ₹142, not ₹95 like you thought.” Expect 3–5% conversion = 75–125 leads from 2,500 group members.
2. Petpooja/Posist integration play: Offer CostKari as a complementary food-cost add-on for Petpooja’s 55K+ restaurant base. Petpooja doesn’t do deep recipe costing — this is additive, not competitive. Pitch a co-marketing deal or marketplace listing.
3. Food blogger / restaurant consultant referrals: India has 200+ active restaurant consultants who charge ₹50K–2L for kitchen audits. CostKari automates the food cost portion of their audit. Offer them a 20% rev-share affiliate commission for referrals.
4. Cold outreach via Zomato/Swiggy listings: Scrape 5,000 independent restaurants from delivery platforms in 4 target cities. Send personalized WhatsApp messages: “Your biryani is listed at ₹350 on Swiggy. After 25% commission and packaging, are you sure you’re making money? Let us show you.”
5. Content play on YouTube/Instagram: Short-form videos in Hindi — “Why your restaurant is busy but broke” — targeting restaurant owner audiences. This is a proven format in Indian business content.

10. Build complexity — justification

Low. The core product is a WhatsApp bot backed by an LLM for bill parsing and recipe calculation, a simple database for recipes/ingredients/prices, and a notification engine for margin alerts. WhatsApp Business API is well-documented. OCR + LLM bill parsing is a solved problem with current models. No custom ML needed. A solo technical founder ships v1 in 4–6 weeks. The hardest part is getting the recipe data entry UX right via conversational flow, which is iteration work, not engineering complexity.

11. Gating checklist

Gate	Pass?	Note
Legal in target market	✅	Standard SaaS. No regulated data. WhatsApp Business API is approved for commercial use in India.
Ethical — no harm / dark patterns	✅	Helps restaurant owners understand their own costs. Net positive.
Market exists (evidence above)	✅	300K+ registered independent restaurants, 65% market share, $7.2B coaching market… wait, food service market at ₹7.76 lakh crore.
1–5 person team can build this	✅	Solo founder for v1. 2–3 people to scale.
Launchable with <$50K / ₹40L	✅	WhatsApp API costs are usage-based (~₹0.50/message). LLM inference under ₹30/customer/month. Cloud hosting ₹5K/mo. Total: under ₹5L to launch and run for 6 months.

12. Feasibility score

Axis	Weight	Score	Notes
Problem intensity	20	17/20	Restaurant owners literally don’t know which dishes lose money. 40–45% food costs vs 30–35% target. ₹3–6L annual waste. This is a hair-on-fire problem for anyone paying attention.
Demand evidence	15	12/15	Strong indirect signals — Toyaja article on ₹3.5L mistake, restaurant consultant industry exists to solve this, POS players adding inventory features. No direct demand for a WhatsApp recipe costing tool yet (it doesn’t exist), but the pain is loudly documented.
Build feasibility	15	14/15	WhatsApp API + LLM + Postgres. Off-the-shelf everything. Solo founder ships in 4–6 weeks. The OCR bill parsing is the only non-trivial piece, and it’s well within current model capabilities.
Distribution clarity	15	12/15	Restaurant WhatsApp groups, Petpooja partnership, consultant referrals, cold WhatsApp outreach from delivery platform listings. Multiple concrete channels. Conversion rates uncertain — ₹499/mo is impulse-buy pricing but restaurant owners are notoriously hard to sell software to.
Revenue mechanics	15	11/15	₹499/mo is the right price for the market. 7,000 restaurants for $1M ARR is achievable but requires serious distribution grind. Churn risk: if the owner stops logging purchases, value drops. Need to nail the habit loop.
Time to first revenue	10	8/10	Can charge from day 1 of WhatsApp access. 7-day free trial → ₹499/mo. First paying customers within 2–4 weeks of launch if distribution channels work.
Defensibility	10	4/10	Low moat initially. Any POS player could add this. Defensibility comes from: (a) recipe data lock-in — once 200 recipes are entered, switching cost is real, (b) supplier price history as a data asset, (c) WhatsApp-native UX that POS players won’t bother replicating because they’re desktop-first.
Total	100	78/100

13. Qualitative modifiers

Founder-fit tags

technical-heavy — Needs a builder who can wire up WhatsApp API + LLM + OCR pipeline and iterate on conversational UX. No sales-heavy founder needed at this price point; the product sells itself or it doesn’t.

Key assumptions to validate (3–5)

1. Assumption: Restaurant owners will consistently log daily purchases via WhatsApp (the habit loop). How to test: 30-day pilot with 20 restaurants. Track daily active usage and drop-off curve. Target: >60% still logging at day 30.
2. Assumption: LLM-based OCR can parse Indian supplier bills (handwritten, mixed Hindi/English) with >90% accuracy. How to test: Collect 200 real supplier bills from 10 restaurants. Run through GPT-4o-mini vision. Measure extraction accuracy.
3. Assumption: ₹499/mo is within willingness-to-pay for independent restaurant owners doing ₹20–80L revenue. How to test: Offer the pilot at ₹499/mo (not free). If 15+ of 20 pilot restaurants pay after trial, pricing is validated.
4. Assumption: Knowing dish-level food cost actually changes owner behavior (repricing, portion adjustment, supplier negotiation). How to test: Track menu price changes and supplier switches in pilot restaurants over 60 days. If <30% take action, the insight alone isn’t enough — need to add more automation.

Risk flags

1. [Habit formation risk]: The product’s value depends on daily purchase logging. If owners stop inputting data after week 2, the copilot goes blind. Mitigation: make input dead simple (photo of bill), send daily reminders, show immediate value (“you saved ₹X today by catching that over-order”).
2. [Platform dependency]: WhatsApp Business API is the sole channel for v1. Meta could change pricing, rate limits, or policies. Mitigation: the web dashboard provides a fallback, and the data/logic layer is platform-independent.
3. [POS player commoditization]: Petpooja or Posist could build this as a feature. Mitigation: they’re desktop-first with bloated UIs. A focused WhatsApp-native tool will out-execute a bolted-on feature. Speed and UX focus are the moat for the first 12 months.

14. Structured verdict

Score:                  78/100
Verdict:                GO
Confidence:             Medium
Best-fit builder:       Technical solo founder with India market experience, comfortable with WhatsApp API and LLM integration
Time to revenue:        4–6 weeks to v1, first paying customers in week 6–8
Capital to launch:      ₹3–5 lakh ($3,500–$6,000)
Top 3 assumptions to validate first:
  1. Daily purchase logging habit sustains beyond 30 days (pilot with 20 restaurants, >60% retention)
  2. LLM OCR parses Indian supplier bills at >90% accuracy (test with 200 real bills)
  3. ₹499/mo converts — 15+ of 20 pilot restaurants pay after 7-day trial
Kill criteria:
  - Abandon if <40% of pilot restaurants are still logging purchases daily after 30 days
  - Abandon if bill parsing accuracy stays below 85% after prompt engineering iterations
  - Abandon if <10 of 20 pilot restaurants convert to paid after trial

15. Next step — 1-week validation sprint

• Day 1: Collect 100 real supplier bills from 5 restaurant owners in one city (offer ₹500 for the stack). Run them through GPT-4o-mini vision API. Measure extraction accuracy for item name, quantity, unit, and price.
• Day 2–3: Build a bare-bones WhatsApp bot (Twilio/Gupshup + Node.js) that accepts a bill photo, extracts items, and confirms with the owner. Add a simple recipe builder: “Tell me what goes into your dal makhani.”
• Day 4: Onboard 10 restaurant owners (personal network + 2 restaurant WhatsApp groups in your city). Ask them to send their daily purchase bills for the next 7 days.
• Day 5–7: Monitor daily engagement. How many send bills each day? Do they ask follow-up questions? Do they share the cost breakdowns with their chef?
• Day 7: Decision: If 7+ of 10 restaurants sent bills on 5+ of 7 days, and bill parsing accuracy is >85%, proceed to build full MVP. If not, diagnose why (too much friction? wrong persona? inaccurate parsing?) and iterate or kill.

The validation must produce a falsifiable result: daily bill submission rate and parsing accuracy over 7 days with 10 real restaurants. Not “people said they liked it.”