AdaptLM

🚀 Inspiration

What’s your favorite LLM to use?
✨ Mine is Claude — I love how well it handles coding.
✨ Sometimes I switch to Groq — it’s just so fast.

But juggling between LLMs means:

💡 No more. That’s why we built AdaptLM.

AdaptLM is an intelligent LLM orchestrator that:

Automatically selects the best LLM for your prompt based on context and task
Seamlessly integrates multiple LLMs in a single session — with shared context and no manual switching
Uses RAG to share context across models, so every LLM stays informed
Routes prompts using Vellum AI benchmarks and advanced NLP, matching intent to the ideal model
Provides agentic tools to handle tasks like emailing and calendar setup — all within the same flow
Supports voice interaction via Vapi, so you can speak directly to your AI assistant

Too busy to type?
✅ Simply enable voice via Vapi and interact with the selected LLM hands-free.

A visual diagram can be represented below:

AdaptLM screenshot

Capturing the semantics and intent of the prompt
Designing an NLP-based router that accurately understands both the task and subtle intent behind a prompt — and maps it to the right LLM — required significant fine-tuning and experimentation.
Addressing the bottleneck of vector embedding ingestion
Ingesting embeddings into a vector database can introduce latency — but users still need instant access to context. We tackled this by designing a solution that keeps the user experience smooth while embeddings process in the background.
Integrating Vapi for voice control
Ensuring smooth speech-to-text and text-to-speech conversion, while maintaining accurate model selection and context continuity, was a complex task involving both frontend and backend coordination.

Innovative solution for dynamic LLM selection
We successfully designed and implemented an architecture that tackles the challenge of choosing the right LLM for the task — without user intervention.
Seamless handling of multiple LLMs with shared context
AdaptLM can switch between models while maintaining context awareness, thanks to our RAG-powered pipeline.
Integrated agentic services without extra steps
Users can send emails, create calendar events, and more — all within the same flow, no manual setup or switching required!

Model selection is harder than it looks
Understanding the true intent of a prompt and matching it to the best LLM requires more than just keywords — it demands careful NLP design and real-world testing.
Context sharing across models is critical
Without seamless context transfer (via RAG), switching between models breaks the user experience. We learned how vital it is to maintain continuity for a smooth workflow.
Voice and agentic integration add huge value — but also complexity
Combining voice interfaces and agentic services with multi-LLM orchestration taught us a lot about managing asynchronous systems and delivering a consistent, responsive UX.

Scale to more LLMs
We plan to expand support to additional models, giving users even greater flexibility and performance options.
Universal API
Our goal is to build a single API layer that lets developers access multiple LLMs through one unified endpoint — no more juggling different APIs.
Define a prompt taxonomy for smarter routing
We aim to create a clear taxonomy for prompt types (e.g. cost-efficient, reasoning-heavy, simple language tasks) to enable faster, smarter, and more cost-effective model selection.