The Role
This is a pure Individual Contributor role within our R&D function. You will work across STT, LLM, and TTS systems - with a clear mandate to close the distance between research and production. That means working directly with backend engineers and systems to ensure your work integrates into live systems cleanly, quickly, and with the observability it needs to be trusted at enterprise scale. The problems are technically deep. The expectation is that they ship.
What You’ll Do
- Lead optimisation work on STT/ASR systems - improving transcription accuracy and reducing latency for domain-specific, multilingual voice data across Indian languages and financial services contexts
- Evaluate, fine-tune, and deploy LLMs for BFSI-specific tasks: information extraction, classification, summarisation, and compliance signal detection
- Build and benchmark TTS capabilities against real product requirements - model quality, naturalness, latency, and integration fit with downstream systems
- Design and maintain scalable prompt engineering and RAG infrastructure for production LLM features
- Work closely with backend engineers and systems (hands-on) to take research outputs from working prototype to deployed, observable production feature
- Establish evaluation frameworks that measure what actually matters - reproducible, deliberate, and tied to real product outcomes
- Track developments in open-source LLMs and ASR frameworks and make reasoned, evidence-backed decisions on adoption
- Identify high-leverage research problems and contribute to where the team invests next
What We’re Looking For
- 8+ years in software engineering with significant depth in ML/NLP systems
- Hands-on experience with LLMs - from prompt design through fine-tuning, evaluation, and deployment
- Exposure to ASR/STT technologies: Whisper, Kaldi, DeepSpeech, or commercial equivalents
- Proficiency with ML tooling: Hugging Face, LangChain, or equivalent frameworks
- Cloud experience (AWS or GCP) for model training, deployment, and monitoring
- Able to make modelling and architecture decisions with incomplete information and articulate the reasoning clearly
- Writes clean, production-ready Python that backend engineers can integrate and maintain
- Understands how ML components fit into larger backend architectures
Strong Signals
- Has closed the gap between “this works in a notebook” and “this is running reliably in production” - more than once, and with an understanding of why that gap exists
- Has worked directly with backend engineers to ship an ML-powered feature and can speak to what that collaboration required technically
- Holds a high bar on evaluation - does not trust a result they cannot reproduce or a metric they did not choose deliberately
- Has made a deliberate build-vs-adopt decision on a core ML component, can articulate the trade-offs, and has lived with the outcome
- Can engage product and business stakeholders on technical constraints without losing precision
Why GreyLabs AI
A hard problem in a large market. Building accurate, low-latency, multilingual Voice AI for regulated financial institutions - across diverse Indian languages and under RBI and IRDAI compliance requirements - is technically complex and commercially consequential.
Real scale, real research problems. The STT, LLM, and TTS challenges here come from actual production load, real customer data, and the constraints of enterprise deployment. They are not synthetic.
Research that ships. At our current stage, the distance between a working experiment and a live product feature is short. Your work will reach millions of conversations.
Strong backing, proven team. Elevation Capital and Z47 are long-term partners invested in our vision. Our founders built and exited Cogno AI - they understand what it takes to build AI companies that earn enterprise trust.