The autonomous procurement agent for cloud kitchens.
Five specialist sub-agents. 9.4 seconds. Daily indent — with auditable reasoning per line.
Synthetic Mumbai cloud-kitchen data. LightGBM quantile forecasts. PuLP MILP optimization. End-to-end pipeline executed in 9.4 seconds.
| SKU | Item | Qty (kg) | Vendor | Cost (₹) |
|---|---|---|---|---|
| SKU001 | Tomato | 41.46 | FreshKart Wholesale | 2,662.98 |
| SKU009 | Capsicum | 12.24 | Mumbai Mandi Direct | 757.17 |
| SKU011 | Cabbage | 7.10 | Maharashtra Agro Hub | 306.86 |
| SKU017 | Beans | 27.58 | Maharashtra Agro Hub | 727.28 |
| SKU031 | Milk | 18.70 | Maharashtra Agro Hub | 1,404.74 |
| SKU049 | Mustard Oil | 15.10 | Maharashtra Agro Hub | 1,080.86 |
| SKU050 | Besan | 11.42 | Maharashtra Agro Hub | 669.33 |
Today's indent commits ₹7,609 across 7 SKUs, led by 41.5kg of Tomato from FreshKart Wholesale (P90 demand 41.5kg). Maharashtra Agro Hub carries the largest share at ₹4,189 (55% of spend), with the remainder split across 2 other vendors. No anomalies detected: all spoilage risks under 50%, no order exceeds 3× P90, and every selected vendor clears the 0.70 reliability floor.
Output above is a real run from the working pipeline. Reproducible: clone repo, python -m fluxon.orchestrator.
The math is structurally hard. Demand, freshness, vendor reliability, and order quantity interact non-linearly, and they change every day with weather, festivals, and supply.
Spreadsheets give up. ERPs track what was ordered, not what should be. Generic AI tools have no domain.
So the buyer guesses. Every morning. Forever.
Domain-tuned, benchmarked against Afresh InvHMM, DingDong Maicai's closed loop, and Amazon Deep Inventory Management.
P10/P50/P90 prediction with weather, festival, and weekend signals.
Knows what's already on the shelf and when it expires.
Routes around the vendor that flaked yesterday. Multi-source allowed.
Minimizes wastage and stockout. Not one or the other.
Flags anomalies. Writes the day's decisions in plain English.
Founder, CEO · Gurugram, India
IIT Delhi (2008). Then nine years running large-scale construction projects — including a ₹210 Cr NRDA Raipur build and the 75-acre Pioneer Park.
In 2019 I joined Grofers (now Blinkit) as Head of Mumbai Operations. In twelve months I scaled the city from 5,200 to 13,000 orders/day — 150,000 line items at peak. Mumbai is India's hardest metro for fresh delivery: density, traffic, real-estate, monsoon supply chain.
I started teaching myself to code in 2021 because hiring tech was strangling the next venture. I'm now CTO at Freshly, a Better-Capital-funded quick-commerce startup, where I built the full stack solo over 18 months — three React Native apps, the orchestration backend, and a 12-component AI procurement system benchmarked against Afresh, DingDong Maicai, and Amazon DIM. Live system: 6,000 orders/month, ₹25 lakh MRR.
Fluxon is the version of that procurement system, built clean-room for everyone else.
Anthropic's Agent Skills (Oct 2025) became an open standard adopted by OpenAI and Microsoft. Claude Managed Agents launched in production April 2026 with Notion, Sentry, and Allianz live. Long-horizon autonomous agents are no longer research demos.
After the 2024 contraction, every cloud kitchen and dark-store operator now optimizes unit economics before growth. Procurement waste — long ignored as 'cost of doing business' — became a CFO-level priority. The category bought what they tolerated for years.
Tom Blomfield's Summer 2026 Request for Startups #3: "AI-native companies that don't sell software — they sell the service." Insurance, accounting, healthcare admin. Fresh procurement is the same shape: a knowledge-heavy service, currently delivered by humans, ripe for autonomous AI delivery.
| Approach | What it does | What it doesn't |
|---|---|---|
| Spreadsheets + buyer | Status quo. 2–4 hours of human guesswork daily. | Doesn't model interactions, doesn't compound. |
| POS / ERP (Petpooja, UrbanPiper) | Tracks what was ordered. | Doesn't decide what should be ordered. |
| Afresh (US grocery) | Distributor-driven supermarket procurement. | Doesn't transfer to mandi-driven Indian supply chains. |
| In-house ML team | Top-3 brands (Rebel, Curefoods) building internally. | 18 months + ₹3 Cr to ship. Locked to one brand. |
| Fluxon AI | Autonomous procurement agent — joint forecast + freshness + vendor + optimization in 9.4s. | Doesn't replace your POS or ERP. Sits beside them. |
What we understand that they don't: fresh procurement is not a forecasting problem. It's a joint optimization problem. Demand, shelf-life, vendor reliability, and order quantity must be solved together — not in sequence. Every other approach breaks them apart. We solve them jointly.
All prices INR, GST extra. Annual contracts -15%.
Afresh is built for US distributor-driven supermarket supply chains. India runs on mandi sourcing — multilingual, weather-volatile, vendor-reliability sensitive — and the SKU mix is different (fresh produce dominant, festivals create 2–3× demand spikes). We rebuilt the math for that.
The top 3 brands probably will — Rebel raised $210M with explicit "integrate AI" thesis. Our long-term ICP is the next 500 brands (10–50 outlets, no in-house ML team). Top brands are credibility / pilot targets, not core customers.
Pilot: anonymized SKU demand only — no customer or vendor PII. Production: SaaS, VPC-isolated, or on-prem container, your CTO picks. We sign DPAs, sit behind your SSO, log every decision for audit.
Yes. Every number on this page is from the actual running pipeline on synthetic Mumbai cloud-kitchen data. The pipeline runs in 9.4 seconds. The indent table above is real output. Reproducible — happy to walk through the code on call.
A fluxon is the smallest unit of magnetic flux — a quantum of flow. Fresh procurement is a continuous flow of demand, supply, freshness, and risk. Naming the company for the smallest unit of that flow felt right.
Live demo. Synthetic data shaped like your category. Thirty minutes. No deck.