Venkat Akhil Lakkapragada

CosmicFish-HRM

Language Model + Research

A language model with a Hierarchical Recurrent Module that learns to allocate compute dynamically at inference time. Pairs with the arXiv publication on adaptive reasoning in compact LLMs.

arXiv Paper GitHub HuggingFace

CosmicFish

Language Model

A modern, efficient language model built from scratch with production-ready deployment capabilities. The flagship model of the CosmicFish family, available at 90M - 300M parameters on HuggingFace.

GitHub HuggingFace

CosmicChat

iOS App · Swift + MLX + CoreML

An iOS application enabling fully offline interaction with CosmicFish language models. Runs models locally using MLX and CoreML on the Apple Neural Engine, with EvoMind moderation integrated to filter harmful prompts. Built entirely in native Swift.

App Store

EvoMind API

Moderation API Platform · Bidirectional Transformer

A content-safety platform powered by a bidirectional transformer, available via API. Five production models: jailbreak detection, profanity detection, PII detection, scam and phishing detection, and harmful content flags. Get an API key and start moderating.

Platform

EvoTalk

Text-to-Speech · PyTorch · Vocos

A single-speaker text-to-speech acoustic model built from scratch in PyTorch. A standalone architecture that predicts mel spectrograms from phonemes, vocoded natively with Vocos at 24 kHz.

GitHub HuggingFace

SpeakScan

Speech + TTS Data

An audio pipeline that downloads YouTube videos, transcribes them, identifies speakers, tags emotions, and generates structured training data for speech and TTS models.

GitHub

EvoMind 2.0

Content Safety

A lightweight 7.4M parameter transformer for real-time content safety classification. Detects harmful prompts, toxic language, and unsafe content. Powers the moderation layer inside CosmicChat.

GitHub

SpecBench

LLM Evaluation Tool

A custom eval suite generator for language models. Describe your model, get a targeted benchmark, score your outputs, and compare architectures and training runs. Built for ML engineers who build small models from scratch.

Live Demo GitHub

Intervo

AI Voice Agent · Sarvam AI Stack

An AI-powered voice interview agent. Upload a resume and job description, conduct a full voice interview, and receive detailed feedback on your performance. Built with Sarvam's speech and language stack.

Live Demo GitHub

PeerX

iOS + Android · Encrypted Messaging

A device-to-device encrypted messaging application for iOS and Android. No accounts, no phone numbers, no server-side message storage.

GitHub

GemmaCare

Flutter · Android App · Health AI · Gemma 4 Hackathon

A personal AI health companion built with Flutter and Gemma. Scans documents, finds nearby care, guides through emergencies, and manages appointments, all with full on-device privacy control.

GitHub

Jericho

RL Environment · Meta Hackathon

A lightweight OpenEnv-compatible reinforcement learning environment where AI agents fix buggy Python code by interacting with a REST API.

GitHub HF Space

Ilyushin

RL Environment · Multi-Agent | GRPO · Meta Hackathon

An OpenEnv-compatible RL environment for adversarial incident escalation and autonomous recovery. A Responder agent trained via GRPO resolves failing services while a Breaker agent generates increasingly deceptive failures.

GitHub HF Space

CosmicFish-Pico

Compact Language Model · 19M Params

A lightweight, efficient language model built from scratch with a custom tokenizer. At 19M parameters, designed for resource-constrained inference.

GitHub

Cosmicizer Tokenizer

SentencePiece Tokenizer

A custom SentencePiece-based tokenizer built for the CosmicFish-Pico language model. Designed and trained from scratch.

GitHub

EvoMind 1.0

Content Moderation · LSTM + Attention

A moderation system that detects and filters messages containing personal information using a hybrid model combining pattern matching with a LSTM model. The first model I built in Dec 2024!

GitHub

EvoMind Mini

Content Moderation · SVM-based

A lightweight version of the EvoMind 1.0 content moderation system using SVM-based classification for detecting and filtering messages containing personal information.

GitHub

TreeCorpus

Dataset · HuggingFace

A comprehensive, structured dataset derived from the latest Wikipedia dumps, processed to serve as high-quality training data for conversational AI models.

HuggingFace

MicroSummaries

Dataset · Mistyoz AI · HuggingFace

A dataset of passages paired with concise summaries, designed for fine-tuning large language models on text summarization tasks.

HuggingFace

CosmicFish-HRM: Adaptive Reasoning via Hierarchical Recurrent Mechanisms in Compact Language Models

arXiv:2605.28919 · 2026

Large language models have achieved strong reasoning capabilities, though often at the cost of massive parameter counts and expensive inference. This work explores a different direction: adaptive reasoning depth in compact language models. CosmicFish-HRM is built around a Hierarchical Reasoning Module that dynamically allocates computational effort during inference, iterating through high-level and low-level reasoning cycles and learning when to halt based on input complexity. The model learns non-uniform reasoning behavior, allocating different numbers of reasoning steps across tasks and inputs, suggesting that adaptive reasoning depth may offer a promising alternative to relying solely on parameter scale.

Cited by IQuest Research (Ubiquant, Beijing) in Loop the Loopies!

arXiv GitHub HuggingFace

NPTEL · IIT Ropar

Deep Learning

Elite · View Certificate

Coursera

Generative AI Fundamentals

View Certificate · Verify

Cisco Networking Academy

CCNA: Switching, Routing, and Wireless Essentials

View Certificate

Cisco Networking Academy

Networking Essentials

View Certificate