Research
We run every model we publish. Our restaurant is our lab: research questions come from real service, systems ship to a live kitchen, and we report what we measured — never what we hoped.
Part of eyeAI Solutions
Curry Creations is the living laboratory of eyeAI Solutions. Every system on this page runs a real restaurant first — real orders, real inventory, real service pressure — and what survives contact with a working kitchen becomes part of ReTech Kitchen, the eyeAI Kitchen OS. Research here is not a demo: it is production, measured daily.
Time-series models on real order history drive what we stock and prep. A recipe-aware procurement engine explodes bills-of-materials against rolling demand, blends statistical usage forecasts with in-flight orders, and gates every price on verification. Forecast error (MAPE) is measured and recorded every day — quality is tracked, never assumed.
LLMs read noisy, multilingual vendor receipts, resolve them to canonical ingredients under strict safety constraints — with a human review gate before anything posts to inventory — and power conversational ordering and generative promotion gated on live orderability. Models are invoked only where language genuinely helps; everything templated rides deterministic queues with a bounded worst case.
Prep graphs and bill-of-materials modeling schedule the work of a one-to-two-person kitchen. Queue-aware pickup ETAs, same-station batch sequencing, coverage analysis that ranks gaps by real order volume, reconciliation of intended work against actual output, and shift check-ins derived from the prep plan itself — so a tiny team runs full service without heroics.
A future direction: activity recognition for workflow and food-safety process monitoring where segmentation removes faces and assigns anonymous track IDs before any analysis. The system observes process — never identities. Not surveillance, by construction.
How small teams and AI systems share work safely: staged-trust rollouts for voice agents, approval-gated autonomy for anything customer-facing, engineered organizational memory for AI collaborators and shift handoffs alike, adversarial multi-agent review of our own systems, and honest, pre-registered evaluation before autonomy.
What the lab shipped and learned lately. Deployed means running in production today; Measured means a number exists and is recorded — we never blur the two.
Two unrelated defect classes each became a ratchet: a contract test that scans the codebase for the defect's fingerprint, fails on any new instance, and carries a baseline that may only shrink. One guards a structured-logging call form that silently drops its context; the other guards test files that name source paths as plain strings, which an import-based reachability check cannot see. The method is the reusable part, and it is independent of either defect.
An unattended self-service payment surface once showed raw processor error text when a card failed. Every failure path now runs through the same failure classifier the staff terminal uses: the customer sees one fixed, plain-language decline sentence, and staff see the classified reason. The classification is computed once and only the rendering differs by audience — there is no second, weaker implementation left to drift.
A nightly sweep runs a registry of cross-cutting data-integrity invariants — business-day-window sanity, cost-versus-revenue plausibility, zero-revenue lines, payment completeness, stuck orders, costing coverage — each check encoding a defect class the platform has seen once. Findings deduplicate on a stable fingerprint and accumulate occurrences instead of piling up rows, and a sweep that fails to run is itself a finding.
On the research portal, a measurement freezes its claim, hypothesis, metric, method, and target before any data is seen — the API rejects edits after the freeze. Findings promote from built to measured only when a number lands, and every review binds to the exact text it read, with diffs naming what changed since.
The receipt-line matcher no longer creates inventory identities from unfamiliar labels. An alias dictionary that once spawned records named after branded package variants now resolves every entry to one canonical ingredient, carrying package size as metadata; an unmatched base name is offered to the reviewer as a suggestion, never created; and a fuzzy match below a fixed confidence threshold is never applied on its own. This closes the double-bookkeeping failure where purchases post to one identity and consumption drains another.
Reassembly emits a proposal — it writes nothing on its own — with every unresolved conflict enumerated for a person. Where one item name survives at two different prices across fragments, the data cannot separate a line read twice from two genuinely identical packs, so that case is handed back rather than resolved by preference. Deduplication is scoped across fragments only: a line the document really printed twice survives.
When one long document is captured as several overlapping photographs, the intuitive stitch — find where one fragment ends and the next begins, then splice — has nothing to work with. On a single production document, position-based splicing found an overlap of zero while 30 of 34 lines matched somewhere across the fragments by content. That is one observation, not a rate. The engine therefore matches on normalized item name plus exact price, and judges the reassembled result against the total the document itself prints rather than rebuilding one from dozens of noisy small-figure reads.
The machinery to identify a customer who has no account was complete and present in the codebase for 69 days before any request could reach it, because a guard written 26 days earlier — encoding a limitation that was real on the day it was written — still stood in front of it. Removing the guard, not building the capability, was the whole of the repair. Systems accumulate guards faster than they retire them, and a guard is hardest to see when the code behind it is correct: nothing fails, the capability is simply never observed to work.
An automated release pipeline was found to contain a promotion stage that could not run at all: it waited on an event that is never delivered for machine-initiated runs, so the stage emitted no output, no error, and no run record. Its only symptom was absence — nothing to read, and nothing that announced there was nothing to read. The repair removed the dependence on cross-stage event delivery entirely, moving the promotion into the same run as the validation it depends on rather than making the missing event more reliable.
For a monetary action an operator previews before confirming, the figure on screen is computed by the same function that writes the record — equal by construction, not by a separately maintained estimate. A later correction is an additive ledger event linked to the original, never a rewrite, and it applies at most once under a guard evaluated inside the writing transaction, so a second submission cannot correct twice.
One canonical low-stock rule now separates "untracked" from "low": an ingredient with no par set is never reported low or out, so never-received, never-used rows stop burying the real shortages. Every consumer routes through the one module, storage-side comparisons that cannot call it carry an explicit positive-par guard, and a source-level fence test enforces both. The flip side — a used ingredient with no par running out silently — becomes a named warning in the nightly invariant sweep, pointing at the surface where the threshold is set.
Per-dish recipe costing is all-or-nothing: if any ingredient cannot be unit-converted and priced from a verified source, the dish reports no cost at all rather than a partial one — because a partial cost understates cost exactly where the data is weakest. A missing number that says so beats a plausible number that lies.
A single owner-confirmed portion rule is the source of truth for a whole family of dishes. The prep-graph mappings, the per-dish deduction recipes, and the unit-converted cost of each dish are all generated from it mechanically, and a daily check re-runs the generator in dry-run mode and flags any dish whose stored recipe has drifted from the rule. Generate, then detect drift: several consuming systems stay consistent without hand-maintaining each one.
Streaming neural speech is infeasible on the kitchen’s constrained hardware — measured real-time factors of 1.6–2.3, plus a twelve-second model load — so the first time a phrase is needed it plays instantly in formant synthesis while a neural rendering is cached in the background. Repeats play the cached audio: latency never regresses, and because kitchen replies are highly repetitive, voice quality converges on the phrases that actually get said.
Best-seller rankings are recomputed nightly from settled sales in one idempotent set-based pass — nothing increments inside payment paths, and cancels and refunds self-correct on the next run. One recompute that is always right beats a counter on every settlement path, where a single missed hook under-counts forever.
One connected system: language models feed inventory, inventory feeds forecasts, forecasts drive the kitchen — and every link reduces the workload of the person cooking.
Guided Kitchen
A kitchen display guides each order through prep, so quality is consistent regardless of who is cooking.
Language Models
LLMs turn unstructured text — receipts, orders, spoken requests — into structured operational data.
Receipt Understanding
Vendor receipts are parsed and resolved to canonical ingredients, keeping inventory truthful without manual data entry.
Live Inventory
Every ingredient tracked in real time — what came in, what each dish consumes, what runs low.
Demand Forecasting
Order history becomes a forecast of what tomorrow needs — the engine of zero-waste stocking.
Prep Planning
Forecasts and bills-of-materials compile into a step-by-step prep schedule sized for a tiny team.
Guided Kitchen
A kitchen display guides each order through prep, so quality is consistent regardless of who is cooking.
Voice Agents
Hands-busy staff talk to the kitchen: safety-gated voice commands with staged trust before any irreversible action.
Privacy-Preserving Vision
Future: cameras that see activity, not identity — faces removed by segmentation, anonymous track IDs only.
Food Safety
Process monitoring and guided steps make food-safety compliance observable and automatic.
Customer Value
The payoff: faster orders, honest ETAs, always-available favorites, and the best food value per dollar.
Working paper · Language Models in Food Operations
A deployed system that resolves messy vendor-receipt lines to canonical ingredients under a hard safety rule — the model may never invent an ingredient — together with an audited safety analysis and a fully specified evaluation protocol.
Working paper · Human-AI Collaboration in Food Service
A case study of how a very small operation maintains a large production platform by engineering explicit organizational memory for AI collaborators — and an honest measurement of where that memory pipeline leaks.
Working paper · Human-AI Collaboration in Food Service
A failure-modes analysis and staged-trust rollout (shadow mode first, destructive authority last) for a kitchen voice agent whose commands can irreversibly change what customers can order — plus a measurement protocol for false destructive activations.
Abstracts are public. Full drafts are available to the team while under active revision — and to collaborators on request via our contact page.
Curry Creations staff can read every paper in full, search the whole corpus, download it as PDF or Word, read any earlier version, and trace a finding back to the commits that implemented it — or start a paper of their own.
Vision
Our goal is a kitchen where the least possible labor produces the best possible food.
01 — Today
AI and machine learning already optimize stocking, accounting, prep, and every task around the plate.
02 — Next
Vision that understands workflow, prep, and food-safety events — where segmentation removes faces and assigns anonymous track IDs before any analysis.
03 — Always
The system studies the process, never the person. No identity, no surveillance — just better food with less work.
