AI engineering & research · commercial & SBIR/STTR · US-performed

We deliver end-to-end AI solutions.

We train the model, adapt an open-source one, build on model APIs, or build the whole system. You get the complete working solution, proven at every milestone. NeurIPS, ICML, Nature Communications, ISBI, and 146K lines of our own code in production.

Start a conversation See the work

Solutions

Bring us the problem.

Describe it in one paragraph. We pick the right method, train a model, adapt an open-source one, build on model APIs, or build the whole system, and put a fixed price and a date on it. You get a working solution, proven at every milestone. Typical projects run 2 to 12 weeks.

The fastest way to start: a two-week AI quality audit at a fixed price. Scope an audit →

YOUR PROBLEM TRAIN A MODEL ADAPT OPEN SOURCE MODEL APIs BUILD THE SYSTEM METHOD · PICKED PER PROBLEM WORKING SOLUTION PROVEN AT EVERY MILESTONE

Four common shapes of the work, not a menu. The method follows the problem.

YOUR SYSTEM 8 LLM JUDGES 2 MODEL FAMILIES SCORE 15 DIMENSIONS BASELINE ACCEPTANCE +28%

Prove your AI works

Independent LLM judges score your system against criteria you can put in a contract. Measured a 28% quality gain with real users on our own product.

Scope an audit →
TARGET BASE MODEL TUNED QUALITY TRAINING →

Get a model that does the job

Custom training with SFT, DPO, or RL, or an open-source model on your own servers. Ours drove a real Boston Dynamics robot at 95% task success.

Scope a training run →
VEHICLE 0.97 VEHICLE 0.91 TRACK 007 0.42 DROPPED

Make your system see

Detection, tracking, and segmentation for drone footage, medical scans, and industrial video. Published at NeurIPS, Sensors, and ISBI.

Scope a vision build →
BUILD CI TESTS ✓ PASS DEPLOY MONITOR KEEP IT RUNNING

Ship it and keep it running

The app, backend, billing, and CI around the model, monitored after launch. Our own product novelhive.ai has run this way since 2025.

Scope a production build →

Work

UAV3D photorealistic aerial scene with five drones over a simulated city

Benchmark · NeurIPS 2024

UAV3D, a public large-scale benchmark for 3D perception from drone swarms

Detection, tracking, and collaborative perception from UAV platforms, in the nuScenes format.

  • 1,000 scenes · 500K RGB images · 3.3M annotated 3D boxes
  • Public code, dataset, and leaderboard
H. Ye · NeurIPS 2024, Datasets and Benchmarks track Project site →   ▶ Demo video
NovelHive: 200,000 words, 26 voices, 7 minutes
Live product

NovelHive, a consumer AI fiction platform on web and iOS

One prompt becomes a 200,000-word novel with a 26-voice audiobook.

Team · live since 2025 novelhive.ai →
Boston Dynamics Spot running our navigation policy, deployment code on screen
Robotics

Sim-to-real navigation on a Boston Dynamics Spot

Navigation policies trained in simulation, deployed zero-shot at 95% task success.

F. Liu · M.S. thesis, 2024 ▶ Watch the robot →
ACD-DETR detection results on UAV imagery
Peer-reviewed

ACD-DETR: small-object detection in drone imagery

H. Ye · Sensors, 2025 Read the paper →
MatchXML architecture diagram
Paper + code

MatchXML: classification over millions of labels

H. Ye · IEEE TKDE, 2024 GitHub →
DeepPhospho workflow figure from Nature Communications
Paper + code

DeepPhospho: deep learning for phosphoproteomics

F. Liu · Nature Communications, 2021 GitHub →
WeakNucleiSeg two-branch architecture with nuclei imagery
Paper + code

Nuclei segmentation from point annotations only

F. Liu · ISBI, 2022 GitHub →
Text-to-image synthesis comparisons on the CUB dataset
Paper + code

Contrastive learning for text-to-image synthesis

H. Ye · BMVC, 2021 GitHub →
Taxonomy diagram of LLM attacks and defenses
Survey

Vulnerabilities and protections in large language models

W. Liu · arXiv, 2024 Read the survey →
APLC-XLNet model architecture
Paper + code

Adaptive label clusters for extreme classification

H. Ye · ICML, 2020 GitHub →
GMMC t-SNE feature clusters
Peer-reviewed

Generative Max-Mahalanobis classifiers

H. Ye · ECML PKDD, 2021 Read the paper →
Adversarial RL feature selection framework
Peer-reviewed

Adversarial RL for unsupervised domain adaptation

H. Ye · WACV, 2021 Read the paper →
Multi-view 3D detection views on the UAV3D benchmark
Under review

PropBEV: temporal query propagation for multi-view 3D detection

H. Ye · under review, 2026 UAV3D benchmark →

Engagement

01

20-minute call

Tell us what you need. If it's not a fit, we'll say so.

02

Fixed-price SOW

Two pages, on your template or ours. For SBIR/STTR we scope to your award's limits.

03

Milestone delivery

We invoice at each milestone, Net 30. All work is done in the United States.

FIXED PRICE · MILESTONE INVOICED · NET 30 M1 · invoice M2 · invoice M3 · invoice done
146Klines of code in production, novelhive.ai
3.3Mannotated 3D boxes, UAV3D benchmark
14publications, listed below
26voices per audiobook, NovelHive

Team

Technical Lead

Weizhen (Frank) Liu

LLM systems & production ML.

Nature Communications · ISBI · Microsoft Research alumnus
Research Lead

Hui Ye, Ph.D.

3D perception & computer vision.

NeurIPS · ICML · IEEE TKDE · Ph.D. CS
Chief Executive Officer

Maya Christine Liu

Contracting authority: contract administration, invoicing, compliance.

Publications

Tell us what's short-handed.

One paragraph from you. We reply with scope, price, and a delivery date.

Goes straight to our team. We reply within one business day.
NOVELA AI

Fixed-price AI engineering. Livermore, California.

Verify

Google Scholar UAV3D paper Nature Communications SAM.gov

Contact

contact@novela-ai.com (678) 997-7619 How we engage
Novela AI LLC UEI (SAM.gov) J6U5C73EBW27 NAICS 541511 · 541512 · 541715 · 541519 Small business · California LLC 2025 Livermore, CA 94551