AI & ML interests

Using AI to remove redundancy across teams, one role at a time.

Recent Activity

Organization Card

Alchemyst AI

Frontier AI for context and compute at the edge.

We build two things: infrastructure that gives AI systems persistent, structured context, and models small enough to run where the data lives.

Website: getalchemystai.com 路 GitHub: github.com/alchemyst-ai


What we work on

Context Layer

A deterministic context backbone for AI applications. Context Layer gives any AI system durable memory and retrieval without vector database sprawl. Graph-based fusion of vector similarity, keyword search, and graph traversal, scoped through composable group hierarchies. Developers use it to add structured recall to agents, copilots, and autonomous systems.

Used by 9,500+ developers across 11+ Fortune 500 pilots and 2 enterprise deployments. #1 upvoted DevTool on PeerList.

API and SDKs: docs.getalchemystai.com

On-device models

We train compact models for speech recognition, content generation, and tool use, optimized for edge hardware: phones, wearables, embedded systems. Three model families ship under the Alchemyst umbrella.

Family Purpose Models
Pratilekha Speech-to-text Indic ASR across 8 languages
C (Content) Content generation Small language models for constrained environments
X (Action) Instruction following and tool use Large action models for autonomous agents

Pratilekha

Our family of speech-to-text models built for on-device Indic ASR.

The name comes from Sanskrit: prati (in response to) + lekha (writing). Speech rendered into text.

Model family

Model Size on disk Target hardware
Pratilekha v1 Tiny 78.4 MB Microcontrollers, ultra-low-power wearables
Pratilekha v1.1 Tiny 78.9 MB Microcontrollers, wearables (improved v1)
Pratilekha Small 469 MB Smartphones, IoT gateways, dev boards

Languages

Bengali 路 Hindi 路 Kannada 路 Malayalam 路 Marathi 路 Odia 路 Tamil 路 Telugu

Trained on curated data from KathBath, KathBath Hard, Common Voice, FLEURS, and IndicTTS, consolidated and cleaned from our Indic ASR datasets.

Design priorities

Edge-first. All models in the family run local inference. Pratilekha Small fits on a 4 GB device alongside a companion LLM. The Tiny variants target hardware profiles where Whisper and Conformer checkpoints cannot run at all.

Indic-native. Training data, evaluation, and architecture decisions target the phonetic and morphological properties of South Asian languages.

Quantization-aware. Models are validated at INT8. We do not recommend INT4 for the Tiny variants; quality drops fast at this parameter scale.

Availability

We will release Pratilekha models through the Alchemyst AI inference API first, then open weights for public download.

Weights will carry a custom non-commercial license. Details will accompany each model card.

Architecture

Architecture details will ship with the model cards when weights are released.


Alchemyst C1

A small language model for content generation in constrained environments. C1 handles structured language output where a full-scale LLM is too heavy: on-device summarization, entity extraction, and context synthesis for the Alchemyst Context Layer pipeline.

Model Size on disk
Alchemyst C1 50.4 MB

The C family targets content generation at the edge. Future C-series models will expand coverage across languages and task types.


Alchemyst X1

A large action model for tool use and function calling, built on UC Berkeley's Gorilla framework. X1 translates natural language instructions into structured API calls, powering autonomous agents that act on external systems.

Model Size on disk
Alchemyst X1 13.8 GB

The X family targets instruction following and action execution. X1 is the foundation model for Alchemyst's agent infrastructure.


Repositories

Repository Type Description
alchemyst-ai/pratilekha-v1-tiny Model Pratilekha v1 Tiny
alchemyst-ai/pratilekha-v1.1-tiny Model Pratilekha v1.1 Tiny
alchemyst-ai/pratilekha-small Model Pratilekha Small
alchemyst-ai/alchemyst-c1 Model Alchemyst C1 - SLM for content generation
alchemyst-ai/alchemyst-x1 Model Alchemyst X1 - LAM for tool use and instruction following
alchemyst-ai/indic-asr-datasets Dataset Consolidated Indic ASR training data across 8 languages

Research

Applied research on context systems and efficient on-device inference.

  • TTFT-context-size tradeoff: Measuring first-token latency scaling across context window sizes

SDKs and integrations

Context Layer SDKs ship in TypeScript, Python, Go, Ruby, and Java, with first-party integrations for LangChain, LlamaIndex, and Vercel AI SDK. We maintain MCP servers and an OpenClaw plugin (361K stars, 3.2M MAU).

Browse the full ecosystem on GitHub.


Contact

For API access, research collaboration, or enterprise deployment: founders@getalchemystai.com

Built in India for the world.

models 0

None public yet

datasets 0

None public yet