Alchemyst Labs
Intelligence
Sovereignty
Own the densest "intelligence per dollar" for sovereign enterprise AI stack that you can trust.
See howPratilekha.
Pratilekha (प्रतिलेख) is a family of lightweight Indic Speech-to-Text models fine-tuned from OpenAI Whisper using LoRA adapters. Designed for edge and mobile inference, they deliver ~20% better accuracy than Whisper-1 on Indic benchmarks with sub-200ms latency.
pratilekha-v0
Lightweight Indic STT using LoRA adapters on Whisper Tiny
- Base
- Whisper Tiny
- Parameters
- 39M
- Downloads
- 1.17k
pratilekha-v0-small
Higher-accuracy Indic STT using LoRA adapters on Whisper Small
- Base
- Whisper Small
- Parameters
- 244M
- Downloads
- 7.89k
bodh
Lightweight Indic language identity detection for VAD-to-STT routing. Frozen Whisper medium encoder with a learned projector and linear classifier — 98.81% accuracy on 7 Indic languages in 30-40ms
- Base
- Whisper Medium
- Parameters
- ~1.5B
- Downloads
- —
Languages Supported
Pratilekha: Lightweight Indic Speech Recognition
A family of lightweight Indic STT models - supports 12 Indian languages with sub-200ms latency on edge devices, achieving ~20% relative improvement over base Whisper on Indic benchmarks.
Research Papers.
Is large context window all you need?
This paper examines the trade-offs between context sizes and latencies, highlighting the need for improved context retrieval strategies that do not bloat query sizes to the concerned Large Language Models.
BrevityBench: Measuring Intelligence per Dollar for Language Models and Agent Harnesses
A benchmark measuring intelligence per dollar across knowledge work, coding, memory retention, and content writing — for both raw models and model-harness combinations — using atomic task checkpoints, hidden evidence-anchored rubrics, and a token-level cost ledger that attributes every dollar spent.
Listen First: Language Identity Detection with Frozen Whisper and a Learned Projector
A lightweight language identity detection system repurposing frozen Whisper encoder representations and a learned speech projector for realtime VAD-to-STT routing. 98.81% accuracy, 98.76% macro F1, 30-40ms latency on NVIDIA L40S.
Applied AI.
Generative UI
The next frontier of interaction.
For the last 50+ years, interfaces have been THE BIGGEST bottleneck for natural human-machine interaction. Interactions with AI systems should feel as natural as extending your hand. Traditional UI does not.
Hybrid Compute
Optimizing intelligence per dollar.
We are working toward AI systems that make deliberate choices about where intelligence runs: in the browser, on local hardware, or through online inference. Our aim is to balance capability, latency, and cost so that each task gets the intelligence it needs without unnecessary compute spend.
Press ⌘K and ask a question. On supported browsers, our assistant can run a local language model through WebLLM, directly in your browser. You can also switch to a free-tier model through OpenRouter within the same interface.
This is one expression of the direction we are exploring: combining local and online inference to optimize intelligence per dollar.
Browser-local inference requires a compatible browser and an initial model download. Search and context-retrieval tools still use network services.
Harness Engineering.
BrowserStonks
Unify your desktop and browser.
A user's digital life is divided into two firmly separated worlds - the desktop and the browser. BrowserStonks bridges this by connecting with your AI harnesses from your browser.
Realtime Harnesses
The next frontier of interaction.
For the last 50+ years, interfaces have been THE BIGGEST bottleneck for natural human-machine interaction. Interactions with AI systems should feel as natural as extending your hand. Traditional UI does not.