ADM's Winter Haven citrus plant distills orange oil into high-value fractions — but deciding what to run, in which still, from which drums lives almost entirely in tribal knowledge. This brief summarizes what we learned on-site, the KPIs the tool must move, and the value case for building it now.
Winter Haven takes cold-pressed orange oil and distills it into fractions of different concentrations — think one pot of soup where red beans, navy beans, and everything in between must be pulled out separately, in the right order, to the right spec. Leftover cuts go back to work-in-process (WIP) as drums on the floor, waiting for a future run that can use them.
That WIP pile has grown to roughly 4,500 lots. Matching drums to a distillation request means checking GC profiles, wet aldehyde, weights, and still capacities — lot by lot, mostly by hand and memory. The plant knows there is money sitting in those drums; what it lacks is a fast, trusted way to decide how to run them down.
Everything the client described rolls up to a single planning chain: pick the WIP, run the still, blend the cuts, and buy only the fresh oil you truly need.
Given a product code and required mass, which WIP lots should feed which still, with how many passes, to hit spec? Today this takes a senior planner 4–8 hours per work order and depends on tribal knowledge. The tool should also answer the reverse question they care about most: "given these random WIP drums, what could we make — and can sales sell it?" That is the direct lever on the $17MM WIP target.
Once fractions exist, what proportions blend to the finished spec? Daniel's internal Python script works but is static — it doesn't learn from new runs. It gives us a validated starting point to build a learning model on. Sign-off stays with QC (GC + sensory panel).
Plan backwards from finished product through the fraction tree: how much fresh cold-pressed oil must Lisa buy, after existing WIP is used first? The constraint is explicit — minimize fresh oil purchases, driven by the excess-WIP problem, not oil price.
Use process data to suggest run settings. Cody's Aspen modeling is already underway internally. Parked by mutual agreement — but a proven Winter Haven tool could later replicate to Monroe, Kalamazoo, and possibly Rodelle.
We walked the current workflow end to end on-site. Five steps, one person, and most of the 4–8 hours goes into step three: evaluating feed candidates one drum at a time.
Check the Distillation Request Sheet for the product code and mass required.
Some product codes have defined parameters and known feed materials ("Orange Tree Map"). Non-standard products don't — and are the biggest pain point.
Search WIP for candidate lots and confirm there is enough mass on hand. Fresh material is easy; judging whether an odd WIP lot fits a run is slow, manual, and where most of the time goes.
Wet aldehyde first, then the GC profile, to select feeds that should deliver the target fractions.
Open JDE and enter the bill of materials and weights — inputs to the still and expected cuts. Some BOMs can be copied from history; others are built from scratch.
Each business objective maps to a measurable KPI with a stated baseline from the on-site visit.
| Business objective | KPI | Baseline today | Target | Owner / problem |
|---|---|---|---|---|
| Reduce work-in-process inventory | WIP $ value on floor | ~4,500 lots | −$17MM by 2026 | Site leadership · P1 + P3 |
| Cut work-order planning time | Hours per distillation WO | 4–8 hrs | < 30 min proposed plan | Planner (Tom) · P1 |
| Remove key-person dependency | # people who can build a WO unaided | Effectively 1 | Any trained planner + tool | Plant management · P1 |
| Improve production accuracy | First-pass runs meeting spec | ~85% | ↑ vs. baseline (fewer reruns) | Production · P1 |
| Minimize fresh oil purchases | Fresh cold-pressed lbs bought per lb of finished product | Not yet baselined | WIP consumed before fresh oil | Procurement (Lisa) · P3 |
| Faster finished-product formulation | Blend iterations to hit spec (proto FPP) | Static model, manual iteration | Learning model, fewer iterations | Formulators / F&F · P2 |
ADM prefers file exports over any live connection into their network — a constraint that also lowers our security exposure. The design assumes a daily manual batchmetrics dump.
Four value streams, anchored to numbers the client gave us on-site — followed by the full calculation ledger, three scenarios, and payback against a $250K build. Figures marked as benchmarks or assumptions will be validated during scoping.
WIP reduction — the client's own 2026 target. The tool's core job is converting slow-moving WIP into sellable fractions. Beyond the one-time cash release, every dollar of WIP that stays off the floor avoids ongoing carrying cost (working capital, storage, handling, quality drift).
Benchmark: at the ASCM/APQC 15–25% carrying-cost rate, $17MM of avoided WIP ≈ $2.6–4.3MM/yr — worked in full in the calculation ledger below.
Per work order, every time. Senior-planner hours shift from drum-by-drum searching to reviewing a ranked plan. The freed capacity goes to the non-standard products and WIP-burndown campaigns that are impossible to plan by hand today.
Illustrative: ~85–95% reduction in planning time per WO; validate WO volume/yr on-site to convert to dollars.
Key-person continuity. A recent multi-week absence of the sole planner was described as catastrophic. Encoding the "black box" into a tool converts an uninsurable operational risk into a trained, documented capability any planner can run.
Hard to price, easy to justify: the client has already experienced the downside scenario.
Fewer reruns and re-samples. ~15% of runs miss target today, triggering re-distillation, blending workarounds, or material going back to WIP — each costing still time, energy, labor, and QC cycles. Even a few points of first-pass improvement compounds across every run.
Illustrative: each avoided rerun saves a full still cycle + changeover; quantify with run counts in phase one.
| Scenario | Carrying rate · attribution | Carrying avoidance | Re-runs | Planning labor | Annual benefit | ROI on $250K |
|---|---|---|---|---|---|---|
| Conservative | 15% · 60% | $1.5M | $0.10M | $0.08M | ≈ $1.7M | ≈ 7x |
| Base | 20% · 70% | $2.4M | $0.20M | $0.11M | ≈ $2.7M | ≈ 11x |
| Upside | 25% · 80% | $3.4M | $0.30M | $0.15M | ≈ $3.9M | ≈ 15x |
Fixed build cost for the Phase 1 WIP planner.
Recurring, before key-person upside or the one-time cash release.
Every scenario clears the investment many times over.
Of realized carrying-cost savings alone covers the full build.
Assumptions to validate before the SOW — these convert benchmark math into ADM's math: annual work-order volume and cost per re-run (still time, energy, labor, QC), the loaded planner rate, the carrying-cost rate ADM finance actually uses (cost of capital + storage + insurance + obsolescence, via Josh / JDE), and the sellable share of current WIP. Replication upside (Monroe, Kalamazoo, Rodelle) is excluded from this case and is pure additional leverage on the same build.
The client's decision-making chain naturally phases the work — and their trust model ("show me the plan and let me compare it to mine") tells us exactly what the MVP must be.
Ingest the daily batchmetrics dump, 3 years of run sheets, and cost extracts. Map old-to-new product codes with site staff. Deliver a browser tool that proposes ranked run plans — lots, weights, still, passes — side by side with the planner's own logic, with reasons shown for every selection.
Build the learning blend model on top of Daniel's validated static script, with QC keeping final sign-off. Add bottom-up planning: finished product → fraction tree → WIP first → minimum fresh oil purchase recommendation for Lisa.
Integrate temperature/pressure data alongside Cody's Aspen work to suggest run settings. Package the proven Winter Haven playbook for Monroe, Kalamazoo, and potentially Rodelle.
Terms and acronyms used throughout this brief — the distillation vocabulary on the left, the systems and business shorthand on the right.