Data services
- Data mining for hard-to-find operational signals.
- Scalable collection pipelines for structured and unstructured data.
- High-accuracy text, audio and vision labeling workflows.
Enterprise AI infrastructure
Data, models and software — built to reach production.
itbanque designs and operates production-grade data, model and software systems that move from prototype to reliable business value.
03
Software
platforms · integration · observability
02
Models
fine-tuning · evaluation · optimization
01
Data
mining · collection · labeling
Language & LLMs
Fine-tuning, evaluation and model integration.
Computer vision
Video understanding and expert labeling.
Speech & audio
Recognition and translation across languages.
Automation
Workflow, browser and marketplace platforms.
What we do
Three layers. One delivery team.
Integrated service layers that support modern AI delivery — engage one, or let us carry the whole stack.
Model services
- Training and optimization from baseline to production.
- Fine-tuning for speech, language and multimodal domains.
- Evaluation harnesses for reliability and deployment readiness.
Software solutions
- Custom product and platform engineering for AI-first teams.
- Integration across data pipelines, APIs and model endpoints.
- Secure enterprise delivery with observability and maintenance.
Selected work
Shipped, measured, in use.
Supervised action recognition for match analysis
Fine-grained movement classification across full-length competitive match video.
100+
action classes
Real-time
inference
Sneakers Radar
Listing, delisting and repricing across StockX, GOAT and Poizon from one live inventory library.
3
marketplaces unified
Auto
price undercutting
Talk2Dom — AI element locator
Natural-language browser automation that turns intent into precise DOM targets.
OpenAI
-compatible API
2
frameworks: Selenium, Playwright
Multilingual speech recognition for media
Fine-tuned Whisper models for five languages, built for subtitle and indexing workflows.
5
languages
<15%
WER, noisy media
Research
We publish what we learn.
Applied work from our delivery programs, written up so others can build on it.
PaperJuly 2025 · Jian Feng
Adapting Whisper for Japanese-to-Chinese Speech Translation on Lightweight Models
Evaluates lightweight fine-tuning strategies that improve Japanese-to-Chinese speech translation using curated subtitle-aligned corpora.
How we work
Move forward, or fall behind.
Infrastructure isn’t the headline, but it decides whether AI progress lasts. Our team combines infrastructure engineering, ML research, linguistics and product execution.
Define outcomes
Every engagement starts with clear outcomes and measurable milestones.
Build hands-on
We work inside your team, from raw input to validated model output.
Evaluate consistently
Experiments scale, evaluations stay consistent, quality stays traceable.
Hand over cleanly
Documentation and delivery standards built for long-term maintainability.
Reply within 24–48 hours