Dinesh Kumar T
Open to AI/ML engineering roles · hybrid preferred

Dinesh Kumar T

AI/ML Engineer · NLP, LLMs & RAG · Speech-to-Text · Chennai, India

I've spent the last 3 years building AI that runs inside live contact-centre calls: intent models that route callers, speech-to-text that keeps up in real time, and LLM/RAG features served behind FastAPI on AWS. I take models from notebook to a Dockerised API, and I'm the one who debugs them when production disagrees.

3 yrsproduction AI for voice bots and contact centres
~15%higher intent accuracy after replacing the old NLU
5STT engines benchmarked for live and post-call audio
~20%faster ETL from MySQL and S3 into Python services

Experience

Data Scientist / AI Engineer

Mar 2024 – Present

Pulse Telesystems · Chennai

NLP and intent classification

  • Designed and shipped a two-stage intent classification API (process first, then intent) that replaced the previous NLU and improved intent accuracy by about 15%.
  • Compared TF-IDF, Logistic Regression, KNN and BERT models with weighted F1 and cross-validation before every release, and combined them with rule-based logic where it was more reliable.
  • Used Sentence-BERT embeddings for semantic matching, so new intents could be added without retraining.
  • Lowered memory use and latency by loading per-process models on demand; served predictions with confidence scores through FastAPI and Flask on AWS EC2, Lambda and S3.

LLMs and RAG

  • Built LLM and RAG pipelines with LangChain on OpenAI GPT and open-source models, with document embeddings in Amazon Aurora PostgreSQL (pgvector).

Speech-to-text and voice AI

  • Built and benchmarked STT pipelines across Whisper, NVIDIA Parakeet-TDT, Vosk and Kaldi, comparing accuracy, latency and resource use for real-time and post-call transcription.
  • Designed a streaming architecture with FastAPI, Kafka, Redis and WebSockets, with VAD, FFmpeg preprocessing and speaker diarization, plus GPU inference, concurrency limits and monitoring for stable operation under load.
  • Wrote Python AGI services on Asterisk for DTMF handling, call routing and IVR flows used by live voice bots, and traced SIP signalling to find root causes of production call issues.

Data engineering and analytics

  • Built ETL pipelines from MySQL and S3 that cut processing time by about 20%, and Power BI dashboards that helped operations improve efficiency by about 10%.

Open-source projects

Each one is tested, runs in Docker, and reports measured results in its README.

Streaming speech-to-text over WebSockets. A WebRTC VAD cuts the audio into utterances and faster-whisper transcribes each one as soon as the speaker stops. Handles 8 kHz phone audio as well as 16 kHz.

  • 8.9%WER, clean audio
  • 11.4%WER, 8 kHz phone line
  • 0.39×real time on CPU
WhisperVADWebSocketFastAPIDocker

Document Q&A with citations on PostgreSQL and pgvector. Local ONNX embeddings, vector, keyword and hybrid retrieval compared on a SQuAD benchmark, and answers from any OpenAI-compatible LLM.

  • 87.5%Recall@5, vector search
  • 69.0%Recall@5, keyword only
  • 61 msper query
RAGpgvectorLangChainPostgreSQLFastAPI

Two-stage intent classification: first the domain, then the intent within it, with a tuned threshold to flag out-of-scope requests. Trained on CLINC150 (150 intents) and served with FastAPI.

  • 92.6%in-scope accuracy
  • 150intents, 10 domains
  • 6 msmedian latency
NLPscikit-learnTF-IDFFastAPIDocker

Data analysis

Skills

ML / Deep Learning
PyTorchTensorFlowscikit-learnHugging FaceBERTSentence-BERT
LLMs & GenAI
OpenAI GPTOpen-source LLMsLangChainRAGpgvectorPrompt engineering
Speech & Voice
WhisperNVIDIA ParakeetVoskKaldiVADDiarizationAsterisk / SIP
Backend & MLOps
PythonFastAPIFlaskWebSocketsKafkaRedisDockerKubernetes
Cloud & Data
AWS EC2 / Lambda / S3Aurora PostgreSQLMySQLSQLPandasPower BI

Education & certifications

Let's talk

I'm looking for AI/ML engineering roles in NLP, LLMs or speech, hybrid or on-site, in Chennai or open to relocating. Email is the fastest way to reach me.