Artificial Intelligence Developer
The Havfun app | West Bengal, India | Today
full-time | hybrid | senior
skills: llm training, fine-tuning, rlhf, pytorch, jax, hugging face transformers, deepspeed, fsdp, quantization, gguf, gptq, awq, pruning, distillation, behavioral data pipelines, dpo, constitutional ai, multilingual nlp, code-switching, indian language model training, blockchain integration, decentralized data systems
About HAVFUN
HAVFUN is building India's first Gen Z social intelligence platform — an AI-native ecosystem where every interaction generates behavioral data, every user action mints a blockchain token, and the underlying AI model learns, adapts, and self-improves in real time. We are not building a chatbot. We are building a proprietary Interactable AI (IAI) — a living behavioral model trained on social, economic, and decision-making data from millions of Gen Z users across India.
Phase 4 of our roadmap deploys a proprietary 2B–7B parameter LLM replacing all third-party AI (Groq, Gemini). This is the role that builds it.
What you will own (All research and development will belong to the company) :
.Architecture and training pipeline for HAVFUN's proprietary IAI LLM — trained on behavioral, social, and predictive interaction data, not generic internet text
.Self-modifying AI system - model identifies weaknesses from user correction patterns and updates weights incrementally without full retraining
.On-device portable LLM deployment (2B–7B parameters) that users can run locally for free — fine-tuned on their own behavioral data
.Hinglish, regional code-mixing, and Indian linguistic pattern training — the model must understand how India's Gen Z actually types and talks
.Behavioral prediction engine: predicting user actions, content virality, and social outcomes before they happen
Integration layer connecting LLM outputs with HAVFUN's blockchain token economy — AI activity mints tokens at ×2 weight
.AI Avatar system: full-device AI companion that operates across all apps with user permission
What we are looking for
.Deep experience in LLM training, fine-tuning, and RLHF — you have actually trained or substantially fine-tuned a language model, not just called an API
.Strong foundation in PyTorch or JAX; familiarity with Hugging Face Transformers, DeepSpeed, or FSDP for distributed training
.Experience with on-device model optimization — quantization (GGUF, GPTQ, AWQ), pruning, distillation for sub-10B parameter deployment
.Understanding of behavioral data pipelines — you know how to turn interaction logs into training signal
.Familiarity with RLHF, DPO, or Constitutional AI approaches for aligning model behavior without catastrophic forgetting
Bonus: experience with multilingual NLP, code-switching corpora, or Indian language model training
Bonus: knowledge of blockchain integration or decentralized data systems
What this is not
This is not a prompt engineering role. This is not a wrapper-around-GPT role. If your AI experience is primarily API calls to OpenAI, this is not the right fit.
Why join now
The work might be paid depending on the profile, insights into funded roadmap to 10M users, part of the founding team and a Phase 4 plan to deploy India's first socially-trained behavioral LLM. The model you build will be the proprietary moat that no competitor can replicate — because the data it trains on only exists inside HAVFUN.
Equity. Hybrid. Direct access to the founder. Real technical ownership.
Apply: official@havefun4u.com
HAVFUN is building India's first Gen Z social intelligence platform — an AI-native ecosystem where every interaction generates behavioral data, every user action mints a blockchain token, and the underlying AI model learns, adapts, and self-improves in real time. We are not building a chatbot. We are building a proprietary Interactable AI (IAI) — a living behavioral model trained on social, economic, and decision-making data from millions of Gen Z users across India.
Phase 4 of our roadmap deploys a proprietary 2B–7B parameter LLM replacing all third-party AI (Groq, Gemini). This is the role that builds it.
What you will own (All research and development will belong to the company) :
.Architecture and training pipeline for HAVFUN's proprietary IAI LLM — trained on behavioral, social, and predictive interaction data, not generic internet text
.Self-modifying AI system - model identifies weaknesses from user correction patterns and updates weights incrementally without full retraining
.On-device portable LLM deployment (2B–7B parameters) that users can run locally for free — fine-tuned on their own behavioral data
.Hinglish, regional code-mixing, and Indian linguistic pattern training — the model must understand how India's Gen Z actually types and talks
.Behavioral prediction engine: predicting user actions, content virality, and social outcomes before they happen
Integration layer connecting LLM outputs with HAVFUN's blockchain token economy — AI activity mints tokens at ×2 weight
.AI Avatar system: full-device AI companion that operates across all apps with user permission
What we are looking for
.Deep experience in LLM training, fine-tuning, and RLHF — you have actually trained or substantially fine-tuned a language model, not just called an API
.Strong foundation in PyTorch or JAX; familiarity with Hugging Face Transformers, DeepSpeed, or FSDP for distributed training
.Experience with on-device model optimization — quantization (GGUF, GPTQ, AWQ), pruning, distillation for sub-10B parameter deployment
.Understanding of behavioral data pipelines — you know how to turn interaction logs into training signal
.Familiarity with RLHF, DPO, or Constitutional AI approaches for aligning model behavior without catastrophic forgetting
Bonus: experience with multilingual NLP, code-switching corpora, or Indian language model training
Bonus: knowledge of blockchain integration or decentralized data systems
What this is not
This is not a prompt engineering role. This is not a wrapper-around-GPT role. If your AI experience is primarily API calls to OpenAI, this is not the right fit.
Why join now
The work might be paid depending on the profile, insights into funded roadmap to 10M users, part of the founding team and a Phase 4 plan to deploy India's first socially-trained behavioral LLM. The model you build will be the proprietary moat that no competitor can replicate — because the data it trains on only exists inside HAVFUN.
Equity. Hybrid. Direct access to the founder. Real technical ownership.
Apply: official@havefun4u.com
Benefits
equity · hybrid
Get new builder jobs daily: