JeevarathinamV
AI / ML Engineer
I turn frontier AI research into products people actually use. Voice agents that feel natural, retrieval systems that find the right answer, language models that stay grounded when it matters.
Bridging ML research and real engineering
I read research papers the way some people read news. When something interesting lands on arXiv, I want to know if it actually works, so I build a small version and find out. That habit is how I ended up fine-tuning voice models, merging two of them with task arithmetic to get something better than either alone, and shipping things I never planned to ship.
Outside of work I'm usually tinkering with something, reading about how a model was trained, or convincing myself one more cup of coffee is a good idea. The best way to collaborate with me is to throw a hard problem at me and bring coffee.
Location
Chennai, India
Education
B.Tech AI & DS
Current Role
AI Engineer Intern
Anand Institute of Higher Technology
B.Tech, Artificial Intelligence & Data Science · CGPA 8.4
Where I've worked
Four internships across AI engineering, data science, and analytics.
AI Engineer Intern · F22 Labs
Dec 2025 – PresentChennai, India
- Authored 95+ technical POC research documents; evaluated 20+ TTS/STT/LLM/OCR models contributing to internal TTS Leaderboard.
- Fine-tuned 3 production TTS models (Kokoro-82M, XTTS-v2, VoxCPM) reducing WER from 60% to 22% and improving NISQA MOS by 18%.
- Fine-tuned LFM2.5-1.2B Instruct LLM on a multi-GPU server and served it via vLLM; deployed full STT + LLM + TTS pipeline into LiveKit as a production voice AI agent.
- Designed production Hybrid RAG architecture (Qdrant dense retrieval + Groq LLM reranking) achieving ~500ms avg / ~800ms p90 latency; implemented GraphRAG over Neo4j.
- Benchmarked Zvec, Qdrant, and Milvus on RAG retrieval accuracy and latency. Zvec fastest with highest recall; published findings on F22 Labs engineering blog.
- Researched Task Arithmetic for TTS model merging. Combined 2 fine-tuned Kokoro models in shared weight space without retraining, achieving 55% listener preference.
Data Science Intern · Shiash Pvt Ltd
Jul 2025 – Nov 2025Chennai, India
- Engineered data pipelines using Pandas/NumPy to preprocess 50+GB datasets, improving training efficiency by 20%.
- Performed systematic hyperparameter tuning using GridSearchCV and RandomizedSearchCV, boosting classification accuracy by 25% on test data.
- Integrated trained models into Flask APIs for real-time inference, reducing latency by 30%.
Data Analytics Intern · UptoSkills
Jan 2025 – Apr 2025Remote
- Built Power BI dashboards for 500+ colleges enabling regional insights and accreditation analysis.
- Automated data preparation with Power Query and Excel, reducing manual reporting effort by 60%.
AI/ML Intern · Arul Technologies Pvt Ltd
Nov 2024 – Dec 2024Chennai, India
- Developed regression model for real estate pricing achieving R² of 0.85+ using NumPy, Pandas, and scikit-learn.
- Executed full ML pipeline from data ingestion through feature engineering, model tuning, and evaluation.
Featured projects
Production AI systems and research spanning Voice AI, LLMs, RAG, and edge inference.
TTS Fine-Tuning & Task Arithmetic Research
Pioneered Task Arithmetic for TTS, combining fine-tuned female-voice and Indian-accent Kokoro models in shared weight space (α=0.6, β=1.0) without retraining, reaching MOS 4.4 and 55% listener preference. Fine-tuned XTTS-v2 reducing WER 58.4% (18.54% to 7.71%). Published 3 models on Hugging Face.
Real-Time Multilingual Translation
Browser-native live speech-to-speech translation across 5+ Indian languages at ~380ms E2E latency, 25+ concurrent listeners per room. Cut cross-lingual TTS latency 83% (650ms→75ms) via 5-provider benchmarking. Multi-room WebSocket architecture with API-key pool rotation and live cost tracking.
GEO Optimizer: AI-Native Content Engine
Multi-stage AI content engine that researches, verifies, and generates citation-optimized articles directly cited by ChatGPT, Claude, Gemini, and Perplexity. 5-stage pipeline: Question Discovery → Source Authority Mapping → Fact Verification → Hub & Spoke Knowledge Map → Article Generation, with hallucination prevention and resume-from-checkpoint.
Offline LLM on Android: Edge AI
On-device LLM inference deploying LFM2.5-1.2B on Android (Poco X3) via llama.cpp + CMake, fully offline with zero internet dependency. Proves edge-AI viability: a quantized LLM running entirely on consumer mobile hardware with no cloud backend.
AI Hoax Buster: Chrome Extension
Browser-integrated NLP Chrome extension for real-time bias and hoax detection with sub-800ms latency on news and web content. Deterministic inference pipelines with chunked processing, label normalization, reproducible scoring, and manifest-compliant extension logic.
MindVault: Secure Knowledge Assistant
Privacy-first RAG assistant for cited, grounded Q&A over your documents with persistent memory. Security-first build: AES-256-GCM encryption at rest, Argon2id password hashing, short-lived JWT with refresh rotation, and Supabase Row-Level Security with per-user vector isolation. Hardened with per-IP rate limiting and automatic failover across providers.
Technical writing
Published engineering articles on Voice AI, RAG, LLM optimization, and edge inference, mostly on the F22 Labs engineering blog.
Reflection Prompting: Why One Prompt Is Not Enough
Explore the power of reflection prompting techniques to improve LLM outputs through iterative refinement and self-correction mechanisms.
Zvec vs Qdrant vs Milvus: Vector Database Comparison for RAG
Benchmarked Zvec, Qdrant, and Milvus on RAG retrieval accuracy and latency across identical testing conditions: indexing speed, query latency, and retrieval consistency. Identified Zvec as the fastest with highest recall accuracy.
What Is TOON and How Does It Reduce AI Token Costs?
Deep dive into TOON (Token-Oriented Object Notation), a data representation format designed to reduce token usage when sending structured data to AI models, cutting costs without sacrificing accuracy.
How to Run Local LLM on an Android Phone
A comprehensive guide to running large language models locally on Android devices for private, on-device AI inference without cloud dependencies.
I Merged Two AI Voice Models With Math And It Actually Worked
Task Arithmetic research for TTS, combining fine-tuned female voice and Indian accent Kokoro models in shared weight space without any retraining. Achieved MOS 4.4 and 55% listener preference vs 27% baseline.
Tools & expertise
The stack I use to research, fine-tune, and ship production AI systems.
Generative AI & LLMs
Voice AI
Frameworks & Libraries
Inference Frameworks
Databases & Vector Stores
Search Technologies and Analytics
Backend & Languages
Cloud & DevOps
CLI Tools
Models & code in the open
Fine-tuned models published on Hugging Face and research code on GitHub.
Hugging Face
@jeevav62
XTTS-v2 fine-tuned on Indian-English. WER reduced 58.4% (18.54% → 7.71%), semantic similarity +12.1%.
Kokoro-82M fine-tuned on 4,358 Indian-English clips. Proper-noun pronunciation improved 3.4 → 8.8 / 10.
VoxCPM LoRA fine-tune for expressive Indian-English speech synthesis with parameter-efficient adapters.
GitHub
@Jeevav62
Top repositories
Certifications
Verified coursework across data science, generative AI, and engineering fundamentals.
Let's build something
Open to AI / ML engineering roles and collaborations. Reach out, I reply fast.