Machine Learning Engineer (Visionary Minds Company)

Titan Holdings; Philippines; Remote; PHP 5,192,000 - 10,384,000; Part time

Job summary

Visionary Minds seeks an experienced ML engineer to build and operate end-to-end voice assistant fine-tuning pipelines, translating transcripts and interviews into personality-consistent models. You will work on voice adaptation, streaming inference, and real-time serving to meet sub-second budgets.

You will evaluate outputs for authenticity, stay in the model’s lane, and iterate retraining or re-prompts based on real conversations with clients.

Qualifications

• 2-3+ years building ML systems that shipped and stayed shipped.
• Real LLM fine-tuning depth - you can discuss LoRA vs. full fine-tune vs. retrieval.

Responsibilities

• Building the fine-tuning pipeline for each persona.
• Training and adapting the voice (TTS) and ASR models for real speech work.
• Owning the real-time loop: streaming ASR, latency, and end-to-end throughput.
• Writing evaluation to judge if the assistant sounds like the person and stays in its lane.
• Running the rollout with timelines and client deployment.
• Improving production models based on real conversations.

Skills

ML systems

LLM fine-tuning

TTS/ASR

Python

PyTorch

CUDA

real-time serving

clear communicating

Tools

PyTorch

Hugging Face

CUDA

Job description

Visionary Minds builds conversational voice assistants that carry a specific person's knowledge, judgment, and way of talking. Instead of a generic chatbot, each assistant is trained to sound and reason like a particular expert - so someone can talk through a hard decision, get grounded advice, or learn from a domain leader as if that person were on the other end of the line. We've built assistants across a few different areas - personal wellbeing and recovery, organizational leadership, and executive decision-making among them - and we're onboarding new clients who each want their own.

Every new client means standing up a new assistant from their material, end to end: the language model and the voice. That's the work.

What you'd be doing

• Building the fine-tuning pipeline for each persona - turning transcripts, writing, talks, and interviews into a model that answers the way that person would. We use a mix (Llama, Mistral, and API models depending on the client), LoRA/QLoRA where it fits and full fine-tunes where it earns it, with RAG when the knowledge moves faster than we want to retrain.
• Training and adapting the voice - the TTS that makes each assistant sound like the actual person, and the ASR that lets it understand whoever's talking to it. This is real speech work: voice adaptation from limited audio, streaming synthesis, and keeping it intelligible across accents and phone-quality audio.
• Owning the real-time loop. These get spoken to, so you'll be living in time-to-first-byte, streaming ASR, and turn-taking - endpointing and barge-in that don't cut people off - against a sub-second budget end to end.
• Writing the eval that says an assistant is ready - not just word-error-rate and benchmark numbers, but 'does this actually sound like the person, does it stay in its lane, does it handle a hard conversation well.' You'll help us build the tooling around that judgment.
• Running the rollout. You'll set and hold the timeline for each new client - what happens, in what order, when it goes live - and flag early when a date's going to slip.
• Improving the ones already in production, off the back of real conversations: what to retrain, re-prompt, re-record, or rethink.

What we're looking for

• 2-3+ years building ML systems that shipped and stayed shipped. You've owned a model in production and dealt with what happens after.
• Real LLM fine-tuning depth - you can talk through LoRA vs. a full fine-tune vs. better retrieval, and why.
• Genuine speech experience: you've trained or fine-tuned TTS or ASR models and you're comfortable with the vocabulary of the field - neural codecs, endpointing, streaming inference, the usual latency traps.
• Strong Python and the training stack (PyTorch, Hugging Face). Comfort near the metal for real-time serving - CUDA, streaming, quantization - is a real plus.
• You can run your own timeline. This is a contract role on a small team; nobody's building your Gantt chart for you.
• You write and explain clearly, including to the clients themselves, who aren't engineers.
• Have the ability to work independently while producing high-quality work, meeting deadlines, and maintaining attention to detail with minimal supervision.

Nice to have, not required experience

• Voice cloning or speaker adaptation from small amounts of audio.
• Making a model sound like a specific person - persona conditioning, character consistency across long conversations.
• Real-time serving optimization: KV-cache management, speculative decoding, streaming codec encode/decode.
• RLHF or preference tuning.
• You have delivered for clients directly, not just handed work over the wall.

The logistics

• This is an independent contract engagement, fully remote, working with a US team.
• Because the team runs on US hours, we need a few hours of daily overlap - realistically your evening, to catch US mornings.