Novela AI · Fixed-Price AI Engineering, Livermore CA

Solutions

Start from whichever problem is yours. Every engagement ends in a fixed-price SOW with named deliverables and a date.

Prove your AI works

“The demo convinced everyone. Now the contract needs a quality number, and there's nothing behind it.”

We build the measurement first: a panel of LLM judges across model families, with divergence checks where they disagree, written up as acceptance criteria you can paste into a contract. When no benchmark exists for your problem, we build one from scratch, like our 500K-image UAV3D benchmark (NeurIPS 2024). On our own product, 8 judges scoring 15 dimensions measured a 28% quality gain with real users.

In 2-4 weeks, you get:

Eval harness, wired into your CI
Scored baseline on your data
Acceptance criteria, contract-ready

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

Get a model that does the job

“The hosted model is almost good enough, and prompt changes have stopped closing the gap.”

We post-train on your data with SFT, DPO, or RL, picking the method by where the base model actually fails. If your data cannot leave your infrastructure, we adapt an open model you can self-host. Policies we trained in simulation ran zero-shot on a physical Boston Dynamics Spot at 95% task success.

In 4-8 weeks, you get:

Tuned weights, owned by you
Training code and data pipeline
Held-out evaluation report

Scope a training run →

Make your system see

“Off-the-shelf detectors fall apart on our footage, and nobody in-house has done perception.”

We train detection, tracking, and segmentation on your aerial, clinical, or industrial data, using detection transformers when small objects or clutter break standard models. The methods are published at NeurIPS 2024, in Sensors, and as an ISBI oral, and you approve the validation split we score against.

In 6-10 weeks, you get:

Trained perception model
Validation study you can cite
Deployable inference pipeline

Scope a perception build →

Ship it and keep it running

“The last contractor emailed a zip file and left. It broke the first week real users hit it.”

We deliver operating software: the RAG or agent layer plus the backend, client, billing, and CI around it, and we stay on through the first weeks of real traffic. We run this stack ourselves: novelhive.ai has been live on it since 2025.

In 6-12 weeks, you get:

Live monitored deployment
Source repo, tests passing
CI gate on every deploy

Scope a production build →

Engagement

Call, SOW, milestones.

20-minute call

Tell us what you need. We will tell you honestly whether we can do it.

Fixed-price SOW

A two-page statement of work, on your template or ours. For SBIR/STTR programs we scope to fit your award's limits.

Milestone delivery

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

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

14 items, 2018 to 2026.

Under review · 2026PropBEV: temporal query propagation for multi-view 3D object detectionH. Ye · first author
Sensors · 2025ACD-DETR: adaptive cross-scale detection transformer for small object detection in UAV imageryH. Ye · author
NeurIPS · 2024UAV3D: a large-scale 3D perception benchmark for unmanned aerial vehiclesH. Ye · first author
IEEE TKDE · 2024MatchXML: efficient text-label matching for extreme multi-label classificationH. Ye · first author
arXiv · 2024Exploring vulnerabilities and protections in large language models: a surveyW. Liu · first author
M.S. thesis · 2024Towards vision and language models aided object navigationW. Liu · Georgia State University
ISBI · 2022Weakly supervised nuclei segmentation via instance learning (oral)W. Liu · co-first author
G3 · 2022RNA subcellular fraction estimation for RNA localization regulationW. Liu · author
Nat. Comms. · 2021DeepPhospho: deep learning accelerates DIA phosphoproteome profilingW. Liu · co-first author
BMVC · 2021Improving text-to-image synthesis using contrastive learningH. Ye · first author
ECML PKDD · 2021Generative Max-Mahalanobis classifiers for image classification and generationH. Ye · author
WACV · 2021Adversarial reinforcement learning for unsupervised domain adaptationH. Ye · author
ICML · 2020Pretrained autoregressive model with adaptive label clusters for extreme classificationH. Ye · first author
Comput. & OR · 2018Quasi-physical quasi-human algorithm for packing equal circlesH. Ye · author

Hand off the problem. We own the AI solution, end to end.

Work

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

NovelHive, a consumer AI fiction platform on web and iOS

Sim-to-real navigation on a Boston Dynamics Spot

ACD-DETR: small-object detection in drone imagery

MatchXML: classification over millions of labels

DeepPhospho: deep learning for phosphoproteomics

Nuclei segmentation from point annotations only

Contrastive learning for text-to-image synthesis

Vulnerabilities and protections in large language models

Adaptive label clusters for extreme classification

Generative Max-Mahalanobis classifiers

Adversarial RL for unsupervised domain adaptation

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