SPECTRA

UAM Poznań · Laniqo · FedCSIS 2026

SPECTRA

Multimodal AI translation designed for e-commerce.

scroll

The problem

Standard MT costs money.

Bad translations don't show errors — they just silently kill conversions.

40%

shoppers abandon pages in wrong language

CSA Research

18%

avg. sales lift from proper localization

Common Sense Advisory

$25B

annual revenue lost to poor translations

Unbabel / Forrester

Three ways MT loses a sale

Trademark liability riskCOMET +0.49

PL source

Farba akrylowa Śnieżka biała 5L

Standard MT

Snow White acrylic paint white 5L

SPECTRA

Śnieżka White Acrylic Paint 5L

Catastrophic context lossCOMET +0.46

PL source

Bateria kuchenna stojąca chrom

Standard MT

Kitchen battery standing chrome

SPECTRA

Chrome deck-mounted kitchen faucet

Absurd word-for-word translationCOMET +0.43

PL source

Zamek metalowy 60 cm

Standard MT

Lock metallic 60 cm

SPECTRA

60 cm metal zipper

SPECTRA sees the image.

5-step synthesis pipeline

01

Text-only

Standard MT baseline — title + attributes only

02

Text + Image

Visual context fed directly to the model

03

Text + Caption

Auto-generated product description as context

04

Caption + Image

Caption and image combined for maximum grounding

05

Synthesis

Best hypothesis selected from all 4 candidates — always ≥ text-only baseline

+ Fine-tuned model

orthogonal gain · both levers stack

Results that speak for themselves.

+1.20

COMET points gained

fine-tuned + synthesis vs text-only baseline

8/8

datasets × models positive

synthesis beats text-only in every config

13,295

products evaluated

ConECT e-commerce dataset

Compute

Built on real infrastructure.

Model fine-tuned and served on UAM supercomputing cluster — fully on-premise, zero external API calls.

UAM HPC ClusterWMI Poznań · gpu_spot partition
1× NVIDIA A100 80 GBTraining & inference · bf16
28K samples · mix SFTConECT + Multi30k · 2 epochs · QLoRA
vLLM · asyncio6 parallel calls / product

Partnership incoming

PCSS · Poznań Supercomputing and Networking Center

pcss.pl ↗

Live demo · real ConECT logs

Czech product catalog→ Polish

4 real products · biggest text-only → synthesis gains

🔧

🇨🇿 Czech source · Tools

Aku rázový utahovák 200Nm NEO Tools 08-600

Wrong word sense
COMET7997+18

Text-only

Klucz udarowy z napędem 200Nm NEO Tools 08-600

Awaiting…

processing…
🎈

🇨🇿 Czech source · Party

TUBAN Helium v plechovce Crazy Hel

Brand name error
COMET8098+18

Text-only

HEL Helium w puszce Crazy Hel

Awaiting…

processing…
🐶

🇨🇿 Czech source · Pet

Konzerva pro psy se zvěřinou Rafi 300 g

Wrong word sense
COMET6398+35

Text-only

Konserwa dla psów z mięsem Rafi 300 g

Awaiting…

processing…
🥮

🇨🇿 Czech source · Food

Maková náplň Bakalland 850 g

Wrong word sense
COMET8498+14

Text-only

Nadzienie makowe Bakalland 850 g

Awaiting…

processing…

Roadmap

Where SPECTRA goes next.

Translation is the entry point. The same pipeline unlocks the full commerce stack.

01E-commerce

Cross-border expansion

Upload a catalog. SPECTRA reads the photos — no more brand names mangled, no more ambiguous words mistranslated. Full assortment live in Czech, Slovak or Romanian in minutes.

02Operations

Customs & logistics

Parcel declaration says "Glasses". Text-only: ambiguous. SPECTRA reads the scan image and writes "Okulary" — not "Szklanki". Correct tariff, no manual review queue.

03Growth

Catalog enrichment

Source title: "Letnia sukienka". SPECTRA sees the photo — red floral, midi length — and generates for the UK: "Red Floral Midi Summer Dress". A searchable listing, not just a translation.

Ready for production · OpenAI-compatible REST API

Let's run a pilot on your
real merchant catalog.

Give us 1,000 active products from your cross-border merchants. We'll translate them three ways — standard MT, our pipeline, human check — and show you the delta in real conversion context.

Run a pilot batch
University of Adam Mickiewicz × LaniqoFedCSIS 2026 — Multi-Candidate Synthesis in MMTGemma 4 26B · fine-tuned on UAM HPC · vLLM · COMET-validated