Our work

Built where operational decisions get difficult.

Signed enterprise planning work, live utility forecasting, and forecasting evaluations on the data businesses actually operate with. Different constraints. A shared decision foundation.

Explore the work Discuss a use case

Pharmaceutical planning Signed enterprise engagement

When the demand plan changes, can production actually follow?

A fast-growing Indian pharmaceutical manufacturer signed UpSynQ for integrated business planning: one connected workflow from the demand plan to a production plan its plants can run.

Client: Fast-growing Indian pharmaceutical manufacturer
Engagement: Signed integrated business-planning programme
Scope: Demand and commercial consensus · central supply planning · production and material planning

The decision

Every month, sales and marketing move the demand plan. The question that matters is not whether the new number is right — it is whether stock, sources, materials and plant lines can deliver it, and what has to change if they can't.

The programme connects demand review, commercial consensus, central requirement planning, material readiness and plant feasibility into one governed decision flow. Each step changes what the next team can commit to. The hard part is keeping those assumptions, quantities, exceptions and approvals consistent from the first forecast to the production plan.

Why it is hard

Not all stock is stock. Quantities in QA hold or close to expiry are visible but can't be counted against demand.
Supply comes in shapes. Batch sizes, minimum order quantities and source-specific lead times turn a requirement into whole batches.
Capacity alone isn't eligibility. A site with spare capacity is no option if the product isn't qualified there.
Materials gate the start date. A batch can't begin until raw and packaging materials are released by quality control.
Rules carry history. Priority tiers, packaging versions, validated lines and staged approvals all decide what is allowed.

See the planning workbench

Decision trail Illustrative · synthetic numbers

Demand & consensusStatistical 50.0k → commercial +20% → consensus 60.0k a month, flagged for the approver
Net requirement57.0k of 70.0k units usable after QA hold and short expiry → 123.0k to plan
Sourcing & materialPrimary plant eligible; material lots released on days 1, 7 and 16
Production planFive 25.0k batches on one validated line, days 1–24
Review & releasePlanner → planning lead → approver; the frozen version goes to the plant

The shape of the workflow the programme is built around, shown with the synthetic case from our planning workbench.

Power utility Live Shadow-tested

Ninety-six forecasts. One operating day to prepare.

Every day, an Indian power utility's system-control desk needs tomorrow's demand for each fifteen-minute block — and then has to turn the gap into compliant market bids before the exchange window closes.

Client: Indian power utility
Desk: System control and power procurement
Cadence: Daily, day-ahead · 96 fifteen-minute blocks

Live: the forecast

Forecasts for all 96 blocks of the next day are delivered every day through a working forecast portal. Each run is versioned, so every number the desk acts on traces back to the run that produced it. Accuracy is tracked block by block and by time of day — the evening transition and the boundary between consecutive days are where the work concentrates.

Shadow-tested: the bid

A forecast is only half the job. The net requirement has to become a pack of day-ahead bids that respects the desk's bid-count limit, block shapes and import corridors. The bid engine builds a feasible pack in seconds, against roughly 45–60 minutes by hand. It has been run against historical days in shadow; live trading trials are the next step.

Where it leads

Demand changes what to bid and what to generate; that changes how much coal to source, move and blend, and what ash must be handled. Forecasting and bidding are the measurable start of one connected architecture.

Explore one operating day

96.8%

Average block-level forecast accuracy across counted live days

Seconds

To a constraint-aware day-ahead bid pack, in shadow tests

96-block outlook
Net position
Bid construction
Desk review

Illustrative The daily path, drawn on a synthetic curve.

Pharmaceutical forecasting Technical evaluation

A zero in the sales file wasn't always zero demand.

A pharmaceutical manufacturer asked a simple question: can a statistical system forecast its demand at least as well as its process — and see things in the data the process can't? The answer was scored once, on three months sealed away from every model.

Client: Pharmaceutical manufacturer · a separate engagement from case 01
Scope: National, depot and stockist forecasts for the same portfolio
Test: Three sealed months, opened once after models and rules were frozen

What the data was hiding

Audit the files. A consolidated sheet disagreed with the period extracts in thousands of cells; the extracts matched the national total, so they were used and every correction exported cell by cell.
Find the censored demand. Hundreds of depot-months showed zero shipments while the stockists they serve kept buying. Those were supply gaps, not missing demand — repaired before any model learned from them.
Respect the breaks. Early depot sheets tracked only part of the portfolio; that was a reporting change, not a demand collapse.
Reconcile the hierarchy. Depot forecasts add up to the national number by construction; stockist sell-out is used as a sensing signal.
Rehearse, then freeze. More than 25 candidate models were tested on nearly two years of history; the champion and its rules were frozen before the vault was opened.
Open the vault once. Three unseen months, scored at the grain the business benchmarks on.

How the Data Foundry works Integrated business planning

Depot zero, stockists buying Illustrative

Synthetic example of censored demand
Month	M1	M2	M3	M4
Depot shipments	410	0	0	620
Stockist purchases	395	210	180	540
Demand used	410	~ repaired	~ repaired	620

A zero during a supply gap is a stockout, not a lack of demand.

National × SKUforecast of record · scored

Depot × SKUreconciled to national

Stockist × SKUsell-out sensing

8.3% WAPE

Forecast error on a sealed three-month national × SKU pharmaceutical holdout, with no systematic bias. Depot and stockist grains are harder and scored separately.

Retail distribution Technical evaluation

Forty depots. A moving product catalogue. A forecast at the grain the planner actually needs.

Working alongside a leading global supply-chain consulting firm, UpSynQ forecast four months ahead for every product code at every depot of a retail distribution business — and scored it on history the models never saw.

Context: Partner-led evaluation with a leading global supply-chain consulting firm
Scope: 40 depots × every design code · more than 40,000 rows · four months ahead
Test: Three rolling held-out windows; accuracy = 1 − WMAPE

What made it hard

A long, thin tail. Most rows carry very little volume — 92% of rows hold 5% of it — so the accuracy that matters is volume-weighted.
Seasons that move. Summer peaks shifted month between years, so seasonality was learned from each product family across all depots, not from one row's short history.
Codes that change. The largest "discontinued" product had most likely been renumbered, with a sister code rising as it fell — one of 56 likely pairs, treated as continuous demand.
Pushes sales can't see. At least one major code sells once a year in a single push and is silent the rest of the time.

How it was tested

Baselines, statistical models, gradient boosting, hierarchical methods, deep learning and pretrained time-series foundation models all ran the same test: learn from the past, forecast four months, score against what happened, on three windows. The method that shipped is an ensemble. In this evaluation, its 80% forecast ranges were tuned on held-out data to catch about 80% of outcomes.

How models earn their place

Same forecast, five rulers

Depot × code, monthly71%63% copying last year
Depot × code, 4-month total79%72% copying last year
Design code81%
Division93%
National95%

Accuracy (1 − WMAPE), mean of three held-out windows; bars start at 50%. Aggregation hides the hard series — name the grain before quoting a number.

One product, two codes Illustrative

Old code New code — forecast as one continuous series.

Across the four

Four engagements. One way of working.

Each case looks different on the surface. Underneath, the same kinds of data, forecasting, constraint and review building blocks support the work — and the hard, domain-specific parts stay domain-specific.

Reusable foundation

Connectors and source mappings
Data-quality profiling — breaks, censoring, gaps
Forecast evaluation on held-out periods
Configurable constraints
Common execution and decision-view patterns

Specific to each engagement

Domain rules and numerical models — expiry and batches in pharmaceutical planning; block grids, bid rules and corridors in power; sparse, changing codes in retail distribution
Decision objectives — cover and cost, cost and risk of serving power, accuracy where the volume is
The customer's operating process — who reviews, who approves, what gets released

What carries forward

Tested software abstractions and configurations make each later engagement faster to set up. Client data and confidential customer rules are not pooled or carried between engagements.

Explore the platform

Also applied

The same foundation has been applied in technical work on consumer-goods channel and promotion planning, and on constrained clinical staff scheduling.

Talk to UpSynQ

Bring us the decision that breaks your planning cycle.

If you run operations or evaluate decision software for a living, bring a real case — the data, the rules and the trade-off you keep making by hand.

Discuss a use case Explore the platform

Or email hello@upsynq.com