Halley AI™ Lab · Research · Prototypes · Experiments

Research prototypes and model experiments.

Explore the experiments, model work, and emerging AI patterns we test before they become production capabilities. For governed deployments, integrations, and healthcare workflows, start with Implementation.

Powered by OpenAI + Halley-AI OSS models on Hugging Face Hugging Face
In the lab Apache-2.0
Research prototypesEmerging AI workflows & evidence patterns.
Evaluation harnessesRepeatable tests for grounded answers.
MLX 20B models4 / 5 / 6-bit, on-device on Apple Silicon.
Lab-to-production handoffs
In the Lab

What we explore.

Focused experiments that help us validate patterns before they become production offerings.

Use-Case Research

Explore emerging AI workflows, constraints, and evidence patterns before committing to a build.

Retrieval Patterns

Test chunking, hybrid search, citations, and evidence transparency across sample corpora.

Interaction Prototypes

Try lightweight UX concepts for assistants, analysts, intake flows, and tool-calling agents.

Model Experiments

Evaluate open models, quantization tradeoffs, prompt strategies, and domain-adaptation ideas.

Agent Patterns

Prototype multi-step reasoning, tool calls, refusal rules, and human-review flows.

Safety & Evals

Build repeatable tests for grounded answers, low-evidence behavior, and escalation quality.

Spotlight

Context-aware answers from any source.

Halley AI connects structured databases and external docs to deliver personalized, context-aware answers. The American Gem Trade Association (AGTA) upgraded their member directory with AI-powered search — enabling natural-language queries and fuzzy matching, so members find the right suppliers faster, even without exact names or terms.

View full case study
agta · member directory / AI search
blue gemstone supplier in Texas
Lone Star Sapphire Co.
Austin, TX · sapphire, topaz
98% match
Topaz & Co. Gemstones
Houston, TX · lab-created & natural
94% match
Emerald Source NYC
matched "blue" → aquamarine, sapphire
synonym
AGTA Fuzzy matching across the member directory — no exact names needed.
Open Models

MLX models for Apple Silicon.

Quantized GPT-OSS 20B models (group-size 32), built for Apple Silicon (M1–M4) with MLX — fast, on-device text generation. Apache-2.0 licensed.

GPT-OSS 20B — 4-bit

Quant: 4-bit (gs32) · MLX

Fastest, lowest memory; great for local prototyping.

GPT-OSS 20B — 5-bit

Quant: 5-bit (gs32) · MLX

Balanced quality/speed; a strong default for many use cases.

GPT-OSS 20B — 6-bit

Quant: 6-bit (gs32) · MLX

Higher quality with modest overhead; still responsive on-device.

Install: pip install mlx-lm · View all at huggingface.co/halley-ai

Handoff

From lab findings to production.

AI Lab is where we test ideas. When a workflow needs secure integration, data boundaries, healthcare controls, or a production launch plan, it moves into the Implementation path.

Research Output

Prototype notes, retrieval findings, model tradeoffs, prompt patterns, and eval results.

Production Design

Architecture, data-access rules, integration points, human handoffs, and risk boundaries.

Implementation Handoff

A clear path from experiment to governed deployment — including HIPAA-capable planning when PHI may be in scope.

Ask about the lab.

Ask about a prototype, model experiment, eval approach, or lab finding. Need a production build? Start with Implementation.