Mar 2026 to present
Senior AI/ML Engineer
UltraSafe AIDubai, UAE. Hybrid from Mumbai
Duplex audio and agent orchestration behind UltraVoice, on a stack where the LLM, TTS and STT are built in-house.
About
I take open models and make them experts.
Photo coming soon
Portrait, 4:5, on a dark ground
Fine-tuning with SFT and LoRA, retrieval that holds up against real data, and voice agents that speak and listen at the same time. Based in Mumbai, working with the Dubai team at UltraSafe AI.
- Based
- Mumbai, India
- Working
- UltraSafe AI, hybrid with the Dubai team
- Since
- September 2024
Experience
Where the work shipped.
Feb 2025 to Feb 2026
AI/ML Engineer
Nimap InfotechMumbai, India
Client-facing delivery of LLM systems for UAE government, US healthcare and enterprise. Expert models, RAG, tool-augmented workflows.
Sep 2024 to Jan 2025
Software Engineer, contract
SoftDigitsRemote
GAN-based image authentication in PyTorch, deployed as a REST API.
Four systems, newest first.
Selected work
UltraVoice
Voice agents that answer real phone calls, on a stack where the LLM, TTS and STT are all owned in-house.
- Architected the duplex audio path: streaming STT in, streaming TTS out, barge-in throughout
- Built the concurrency and back-off layer so calls degrade gracefully under TTS rate limits
- Owned agent orchestration: state, tool calls, retries, mid-call recovery
Streaming STTNeural TTSWebSocketsPythonTelephony
Try the idea: press and hold, and the agent yields the line to you
Not started
Patient report retrieval
A RAG pipeline giving clinicians patient records in seconds.
- Retrieval across structured and unstructured patient data
- Semantic search by patient name, ID or mobile number
- Backend validation over LLM output for access-controlled medical data
RAGEmbeddingsVector searchPostgreSQLDjango REST
Dubai Government service classification
A fine-tuned expert model routing government service requests to the right department.
- Instruction datasets built from labelled UAE government service descriptions
- SFT with PEFT and LoRA for task-specific domain adaptation
- Constrained inference prompts for deterministic, high-precision classification
SFTPEFT / LoRADataset preparationConstrained inference
Image-based authentication
A GAN-based authentication model, trained in PyTorch and deployed behind a REST API for external client integration.
GANPyTorchREST API
The stack
What the work
runs on.
Models and agents
Retrieval and data
Voice
Serving and backend
Achievements
9,227 TEAMS
District Winner, Maharashtra Student Innovation Challenge
Raigad district. Selected from 9,227 teams by the Maharashtra State Innovation Society.
800+ PARTICIPANTS
Winner, Technoscope 2024
First place among more than eight hundred participants.
Education
B.E. Computer Engineering
2021 to 2025Anjuman-I-Islam's Kalsekar Technical Campus, Navi Mumbai. CGPA 7.52 of 10.
Have a hardmodel, voice, orretrieval problem?
I answer email fastest.