Reads from
AVEVA PI System · Siemens Opcenter · Aspen InfoPlus.21 · Ignition by Inductive Automation · custom historian and CSV exports
Data security
Process data stays in your environment. On-premises and private-cloud deployments supported.
Product pillars
Earlier visibility only matters if it changes what process, quality, and manufacturing teams do next.
Flag elevated defect risk before downstream confirmation arrives.
Example: Surface elevated risk on a lot from early cycles — built on the same class of early-prediction methods that achieve 9.1% test error on the MIT-Stanford-SLAC public benchmark (Severson et al., 2019).
Narrow likely upstream contributors faster for engineering teams.
Example: Rank the likely upstream contributors to a defect family — drying parameter variance, coating thickness drift, formation profile — so investigations open with structure, not a blank screen.
Bring plant history together without pretending the stack is clean.
Example: Stitch MES, historian, and shift-handover sources into a single lot view — even when the data is messy, partial, or fragmented across systems.
Turn investigations into faster stabilization and stronger plant memory.
Example: Each closed investigation feeds back into the model — turning plant memory into compounding intelligence rather than knowledge that walks out the door at shift change.
Operating flow
Lychee shortens the path from fragmented data to earlier engineering action.
Process history
Mixing, coating, drying, calendaring
Lot context
Line, recipe, shift, genealogy
Quality outcomes
Inspection, yield, formation signals
Engineer action
Ranked risk and likely drivers
Example workflow panel
Interface shown reflects pilot deployment structure. Data is illustrative.
Sample risk queue
conceptual interface
LOT-0427
Formation outlier risk
LOT-0419
Coating drift
LOT-0398
Drying variance
Drift flag
Coating thickness deviation
Example confidence
0.84
Recommended investigation
Check slurry feed oscillation and recent slot-die settings.
Likely upstream drivers
Engineer feedback
Architecture
| Dimension | Plant reality | Lychee approach |
|---|---|---|
| Input regime | Sparse defect labels, slow feedback, drift across lines and chemistries | Hybrid inference combining battery process knowledge with machine learning trained on plant data |
| Inference behavior | Process drift, lot-level patterns, equipment-fingerprint signals | Ranked drivers with calibrated confidence, not unweighted anomaly scores |
| Output shape | Plant teams need ranked actions, not floating scores | Risk queues, likely upstream drivers, investigation priorities tied to process windows |
| Engineering interaction | Outputs must reach existing operating interfaces | Returns to plant team's existing review and quality systems |
Policy relevance
EU Battery Regulation traceability and digital battery passport obligations from 2027 (chapter VII) require richer per-cell process and lifecycle data than most plants currently capture. Lychee makes that data usable for plant teams first, and auditable for compliance second — a byproduct of better operating visibility, not a separate workstream.
For UK gigafactory programs aligned with the Faraday Battery Challenge and the Automotive Transformation Fund, yield improvement at scale is both an economic and a strategic manufacturing-competitiveness goal.
Used for the 70-80% early-ramp scrap framing.
Used for the $4M/day 50 GWh lost-production reference.
Used for the annual value framing of a 1% yield improvement.
Next step
Move from product narrative to a defined first deployment scope.
Scope A PilotReview the battery-process and hybrid-modeling rationale.
See Product ArchitectureExplore fit with a broader plant or ecosystem context.
Discuss Strategic Deployment
