SightlineOS

SightlineOS is an AI-native supply-chain planning and management platform for restaurant chains — “forecasting, inventory optimization, and COGS management” built for foodservice (story). The technical core is a machine-learning forecasting engine that “captures daily ordering cadence, seasonality, holidays, and recent volume shifts while remaining robust to sparse and volatile data” (forecasting) and predicts the ramp-up for brand-new restaurant openings — then feeds one connected data model that turns demand into 12-week supply-continuity risk and into COGS variance attribution. It came out of stealth in May 2026 after a year in private beta.

Vitals: founded ~2024 · seed-stage (amount undisclosed) · ~5–10 people · New York.

Low public engineering disclosure — read the tiers

SightlineOS is a ~5-person, just-out-of-stealth company with no engineering blog and no public job board. Product behaviour, founders, and customer results are well-verified; the engineering internals (languages, cloud, model architecture) are labeled inference from founder pedigree and product behaviour, not first-party fact. The Stack table is intentionally thin; the reasoning lives in Likely internals.

Business context — founders, funding, customers, traction

• Founders: Yusha Hu (CEO) led supply-chain & procurement teams at Chipotle, sweetgreen, and HelloFresh; Derrick Staten (CTO) was engineer and head of product at Branch (“one of the world’s most widely used mobile marketing platforms,” clients incl. Starbucks, Uber); Louis Bensard (Head of Data/AI/ML) was a data-science manager at Ekimetrics, building ML for Fortune 500s in retail/finance/aviation (About). Plus Jake Anderson (Sales, ex-Olo/Cardlytics/Groupon) and Emily Schultz (Marketing, ex-Clover/BentoBox).
• Funding: seed-stage; no figure disclosed first-party. (A widely-repeated ”~$2M seed, Feb 2026” traces to a recruiter posting that doesn’t name the company — treat as unconfirmed.)
• Customers: Din Tai Fung (highest US AUV, ”+$27MM average AUV per location”) and Bonchon (case study, home).
• Din Tai Fung results: “14% lift in forecast accuracy”; within five months a “25%” cut in priority-SKU distributor out-of-stocks (13% across all proprietary SKUs) and a “99.7% fill rate” (case study). Bonchon: sub-3-month, “hands-off” rollout (home).
• Launched out of stealth May 6 2026 after a year in private beta (press).

The heavy lifting

• Forecasting that survives sparse, spiky restaurant data. The ML engine “captures daily ordering cadence, seasonality, holidays, and recent volume shifts while remaining robust to sparse and volatile data” (forecasting) — restaurant ingredient demand is intermittent and bursty (LTOs, local events), exactly where naive per-SKU time-series break; the bet is models that stay stable when history is thin.
• Cold-starting demand for stores that don’t exist yet. It “predict[s] the ramp-up for new restaurant openings — so your supply chain is ready before the doors open,” folding new openings, LTOs and campaigns into the forecast from analogous history (forecasting) — beating the formula-based legacy tools that have no signal for a SKU or site with zero sales.
• One connected data model, not three reports. Spend is wired “directly to your inventory and forecasting,” so a COGS drift can be attributed to a “pricing issue, a forecasting miss, or a supplier problem” rather than read off disconnected variance reports (cogs) — the join across forecast/inventory/spend is the product, not a dashboard.
• Risk detection 12 weeks out at item × distributor. It “identifies supply continuity risk up to 12 weeks in advance by looking at open POs, supplier lead times, and expected depletion” and “sets optimum inventory levels at the item x distributor level” (inventory) — turning forecasts into a forward-looking, per-supplier action queue instead of a retrospective stock view.

Stack

Deliberately thin — only what’s publicly evidenced. SightlineOS names no languages, frameworks, cloud, or model libraries; the rest is in Likely internals.

Layer	Choice	Evidence
Product	web SaaS platform: forecasting + inventory + COGS modules, dashboards, scorecards	home, inventory
ML	in-house demand-forecasting engine (per-customer, continuously retrained on owned data)	forecasting, About
Inbound integration	SynergySuite — bidirectional sales/operations ↔ supply-chain data	press
Data ingested	distributor invoices, open POs, supplier lead times; 700+ commodity market feeds	inventory, cogs
Outputs	demand forecasts; 12-week risk alerts; supplier/DC scorecards; invoice reconciliation; dead/excess-stock flags	inventory, cogs

Key finding — the forecasting engine is the moat, and it’s an applied-ML problem, not an LLM one

SightlineOS is built around classical demand forecasting on the customer’s “owned historical data,” tuned to be “robust to sparse and volatile data” (forecasting) — led by a founder (Louis Bensard) whose background is applied ML at Ekimetrics, not generative AI. The defensible asset is forecast accuracy on hard restaurant data (the “14% lift” at Din Tai Fung), not a chat interface.

Hard problems

The parts an engineer here loses sleep over. Public signal is cited (verified); likely approach is labeled speculation — best-practice fill-in, hedged.

Problem	Why it’s hard	Public signal	Likely approach (speculative)
Sparse, volatile per-SKU demand	Ingredient demand at item × location × day is intermittent and spiky (LTOs, weather, local events); naive per-series models overfit noise	engine “robust to sparse and volatile data” capturing “daily ordering cadence, seasonality, holidays … volume shifts” (fcast)	A global/hierarchical model (e.g. gradient-boosted trees over pooled features) that borrows strength across SKUs and sites, plus intermittent-demand handling — not one ARIMA per SKU
Cold-start: new stores & SKUs	A new opening or LTO has zero history, yet must be ordered for before launch	”predict the ramp-up for new restaurant openings”; incorporate LTOs/campaigns from historical data (fcast)	Analog/pooled priors — borrow ramp curves from comparable openings & similar SKUs; blend into the forecast as actuals arrive
Forecast → inventory → COGS coherence	The three modules must agree on one truth so variance can be attributed, not just reported	spend wired “directly to your inventory and forecasting”; COGS drift traced to pricing vs forecast vs supplier (cogs)	A shared semantic data model (item, location, distributor, contract) under all three surfaces; attribution joins over it
Messy distributor/supplier data	Invoices, POs and distributor buying systems differ per partner; item × distributor needs clean entity resolution; 700+ commodity feeds	item × distributor levels; invoice reconciliation; data that “couldn’t keep pace” on legacy tools (inv, dtf)	ETL + entity resolution — SKU/UOM normalization across distributors, contract-price repository, commodity-feed joins

Likely internals

What SightlineOS doesn’t disclose, inferred from product behaviour + founder pedigree (Branch web/product eng; Ekimetrics applied ML). All speculative/inferred — flagged, not fact.

Component	Likely choice	Basis
Forecasting models	Python; gradient-boosted trees / global hierarchical time-series with intermittent-demand handling; reconciliation item→category→network	”robust to sparse and volatile data”, item×distributor, network-wide (fcastinv); Ekimetrics applied-ML pedigree (About); libraries unnamed
Cold-start	pooled/analog priors + ramp curves for new sites & SKUs	”predict the ramp-up for new restaurant openings” (fcast)
Web app	TypeScript / React (likely Next.js) front end, Node API	Branch consumer-scale web/product pedigree (About); modern SaaS dashboard; specifics unstated (a recruiter JD suggests this but doesn’t name the company)
Cloud	AWS	conventional for a NYC seed SaaS; not stated
Datastores	Postgres (relational core) + Redis (cache/jobs); a columnar/warehouse layer for analytics	spend/inventory/forecast joins, real-time spend reporting (cogs); engines unnamed
Data ingestion	batch ETL + entity resolution; SynergySuite API; commodity-data vendor for 700+ feeds	SynergySuite bidirectional (pr); 700+ commodities (cogs); vendor unnamed
Tenancy	per-customer models trained on each chain’s owned data	”continuously learns from your owned historical data” (fcast)
Funding / stage	seed; figure undisclosed	no first-party number; the ”~$2M” figure is from an unattributed recruiter post

Architecture

Platform: many feeds, one data model, three surfaces

The reconstruction below is verified at the edges (sources, product surfaces, integration) and inferred in the middle (the normalization + unified-model layer). Sales/POS data arrives bidirectionally through SynergySuite; distributor invoices, open POs and supplier lead times, the chain’s owned demand history, and 700+ commodity market feeds join it. That feeds the ML forecasting engine and a connected data model under all three product surfaces — Forecasting, Inventory Optimization (item × distributor, 12-week risk), and COGS — which surface to the supply-chain team as dashboards, risk alerts and supplier scorecards (forecasting, inventory, cogs, press).

Mermaid source

flowchart LR
  classDef src fill:#eef2f8,stroke:#94a3b8,stroke-width:1.5px,color:#0f172a;
  classDef data fill:#e8f1fd,stroke:#2563eb,stroke-width:1.5px,color:#0f172a;
  classDef ai fill:#eafbf1,stroke:#16a34a,stroke-width:1.5px,color:#0f172a;
  classDef io fill:#fdf4e8,stroke:#d97706,stroke-width:1.5px,color:#0f172a;

  subgraph Sources["Data sources"]
    direction TB
    Sales("Sales / POS<br/>via SynergySuite (bi-directional)"):::src
    Dist("Distributor data<br/>invoices · open POs · lead times"):::src
    Hist("Customer's owned history<br/>per-SKU demand"):::src
    Comm("Commodity markets<br/>700+ (corn · soy …)"):::src
  end

  Ingest("Ingestion + normalization<br/>entity resolution · SKU / UOM mapping"):::data
  Model[("Unified supply-chain data model<br/>spend ↔ inventory ↔ forecast")]:::data
  ML("ML forecasting engine<br/>item × location × day<br/>robust to sparse/volatile data · new-opening ramp"):::ai

  subgraph Products["Product surfaces"]
    direction TB
    Fcast("Forecasting<br/>ingredient demand"):::ai
    Invy("Inventory optimization<br/>item × distributor · 12-wk risk"):::ai
    Cogs("COGS management<br/>commodity tracker · invoice reconciliation"):::ai
  end

  Team(["Supply-chain team<br/>dashboards · risk alerts · scorecards"]):::io

  Sources --> Ingest --> Model
  Model --> ML
  ML --> Fcast
  Model --> Invy
  Model --> Cogs
  Fcast --> Invy
  Products --> Team

Forecasting approach (inferred)

How the engine likely stays accurate on hard data. This diagram is mostly inference (purple nodes) anchored to two verified facts: the data signals it consumes and the 14% accuracy lift it produced. The plausible shape is a global/hierarchical model pooling across SKUs and locations, a separate cold-start path for new openings/SKUs, and hierarchical reconciliation so item-level forecasts sum coherently up the network.

Mermaid source

flowchart LR
  classDef data fill:#e8f1fd,stroke:#2563eb,stroke-width:1.5px,color:#0f172a;
  classDef ai fill:#eafbf1,stroke:#16a34a,stroke-width:1.5px,color:#0f172a;
  classDef spec fill:#f3eefe,stroke:#7c3aed,stroke-width:1.5px,color:#0f172a;
  classDef io fill:#fdf4e8,stroke:#d97706,stroke-width:1.5px,color:#0f172a;

  Hist[("Owned history<br/>SKU × location × day")]:::data
  Cal("Calendar signals<br/>seasonality · holidays · LTOs · campaigns"):::data
  New("New-opening / new-SKU<br/>(no history)"):::data

  Global("Global / hierarchical model<br/>(inferred) — pools across SKUs & locations"):::spec
  Cold("Cold-start<br/>(inferred) — analog priors + ramp curves"):::spec
  Recon("Hierarchical reconciliation<br/>(inferred) — item → category → network"):::spec

  Out(["Stable per-SKU forecast<br/>+14% accuracy at Din Tai Fung"]):::io

  Hist --> Global
  Cal --> Global
  New --> Cold
  Global --> Recon
  Cold --> Recon
  Recon --> Out

Speculation — why classical ML, not an LLM

The product is point forecasts on numeric, sparse, seasonal series with hard cold-start — squarely the home turf of gradient-boosted trees and hierarchical time-series, not language models. Louis Bensard’s Ekimetrics background (econometric/applied ML) points the same way. Likely a GBM-or-global-model core with analog-based cold-start; the framework is unstated.

Team & process

A ~5-person, NYC, supply-chain-operator-led team pairing domain depth with Silicon Valley engineering (About, story).

Role	Person	Source
Co-founder / CEO	Yusha Hu (ex-Chipotle, sweetgreen, HelloFresh)	About
Co-founder / CTO	Derrick Staten (ex-Branch eng & head of product)	About
Co-founder / Head of Data, AI & ML	Louis Bensard (ex-Ekimetrics DS manager)	About

The shape is a classic three-founder split — domain (Hu), product/eng (Staten), ML (Bensard) — with sales and marketing leads attached early. Process signal is sparse by design (no eng blog), but the customer evidence is telling: a white-glove implementation team lands enterprise chains in under three months with a “hands-off” experience for the customer (home, inventory), and the forecasting engine is explicitly built to “get smarter the longer they run” (About) — i.e. per-customer models that improve on each chain’s accumulating data. For an early team selling into operationally exacting brands like Din Tai Fung, implementation speed and forecast accuracy are the whole game.

Sources

Reconstructed from public sources only — no insider information. Crawled 2026-06-10 via Chrome MCP (logged-out). First-party (sightlineos.com — homepage, About, the forecasting/inventory/COGS pages, the Din Tai Fung case study, the founder-story blog post) prioritized; the PR Newswire launch release and SynergySuite note are third-party. Claim tiers: verified (stated on a public page, linked) · inferred (reasoned from a cited signal) · speculative (best-practice fill-in, labeled). This is a low-disclosure, early-stage company: engineering internals are inference, not fact. Links are live; pages change, so the supporting quote for each claim is kept in this repo’s evidence map (evidence/sightline-os-evidence-map.md).

#	Source	Link
S1	Homepage	https://www.sightlineos.com/
S2	About us	https://www.sightlineos.com/about-us
S3	Forecasting	https://www.sightlineos.com/forecasting
S4	Inventory optimization	https://www.sightlineos.com/inventory-optimization
S5	COGS management	https://www.sightlineos.com/cogs-management
S6	Case study — Din Tai Fung	https://www.sightlineos.com/case-studies/din-tai-fung
S7	Blog — Our Story (founder note)	https://www.sightlineos.com/blog
S8	Press — launches out of stealth (PR Newswire)	https://www.prnewswire.com/news-releases/sightline-os-launches-out-of-stealth-bringing-the-next-generation-of-ai-powered-supply-chain-management-and-planning-software-to-enterprise-restaurant-teams-302764001.html
S9	SynergySuite × SightlineOS partnership	https://www.synergysuite.com/blog/synergysuite-sightline-os-announce-integrative-partnership/