Himanshu Panwar profile photo 1 on Tal

Himanshu Panwar

Machine Learning | Ex Observe.AI | AI Agent | Speech AI | Building ASR | Conversational AI

Bengaluru

5+ years of work experience

VerbaFlo.AI
Dec 20259 mosPresent

Researched and deployed task-specific Small Language Models (SLMs), improving reliability, accuracy, and latency, built an intent detection pipeline that reduced latency from 18s to 6s with significant cost savings. Integrated multiple LLM providers (Cerebras, Groq, Gemini, Bedrock) and developed an in-house cross-model prompt benchmarking and evaluation framework, resolving 800+ tickets and remediating 300+ prompts flagged by the benchmark system. Architected a centralized prompt repository with configuration management, fallback logic, validation layers, and automated update pipelines, streamlining prompt operations and accelerating customer escalation resolution.
Observe.AI
Nov 20241 yr 2 mosDec 2025

Bengaluru

Built a real-time transcription enhancement Python library for transcript correction, optimized for high-throughput production traffic with ~3ms per-call latency. Led and delivered the Multilingual Self-Serve Transcription Correction feature in the Observe.AI core product, enabling users to correct transcripts directly from the dashboard across all supported languages, significantly reducing transcription-related escalations. Implemented LLM-based transcription enhancement engine to generate correction suggestions at scale for bulk data, improving transcript quality and reducing manual effort for large datasets.
Observe.AI
Jun 20222 yrs 5 mosOct 2024

Banglore

Built and productionized large-scale ASR systems, iteratively improving in-house models through multi-domain data ingestion, evaluation pipelines, and data cleaning to boost accuracy. Trained large-scale deep learning models using distributed training across A100, V100, and T4 GPUs on AWS and Oracle Cloud using PyTorch DDP, Horovod, and Slurm. Optimized models and inference pipelines using ONNX, Triton, TensorRT, and ESPNET_ONNX, achieving ~75% memory reduction and improved latency in production. Revamped speaker diarization for contact-center mono calls, enabling agent–customer separation, customer-specific customization, and R&D for multilingual end-to-end diarization. Developed an internal speech evaluation toolkit supporting ASR and diarization accuracy evaluation, transcription parsing, normalization, multi-language support, and other functionalities. Developed a customer onboarding and support automation platform deployed across multiple clusters, providing customer-specific analysis, recommendations, enhancements, and evaluation capabilities. Owned and maintained multiple speech services, handling production reliability, monitoring, Python upgrades, and on-call incident resolution with customer-facing impact.
Observe.AI
Sep 202110 mosJun 2022

Led R&D of end-to-end ASR systems using ESPnet (Transformer, Conformer, RNN-T, Conformer-Transducer), classical models in Kaldi and self-supervised learning techniques; Successfully trained world-class acoustic models for ASR using large-scale in-house data for contact center. Trained large-scale E2E-ASR models for production on SageMaker, EC2 Ubuntu instances, and Oracle Cloud. Generated over 100K+ hours of speech training data for ASR, including both golden and silver labeled datasets. Developed domain-specific golden labeled test data to rigorously assess the robustness of our in-house ASR system. Designed and deployed Voice Activity Detection (VAD), reducing audio processing by ~40%, doubling throughput, and significantly lowering production cost. Developed and trained models using the traditional speech recognition system Kaldi. I have hands-on experience with ESPnet, an open-source ASR toolkit.
Mando
Jun 20213 mosAug 2021

C++ code optimization. Code profiling.
training and placement cell, iiit bhopal
May 20211 yr 1 moMay 2022

Maulana Azad National Institute of Technology
May 20213 mosJul 2021

Implemented Cluster-based approach for finding the correlation between the dataset.  Performed multiple experiment i.e. distance correlation, differential privacy algorithms, MIC and calculated information loss.  Developed Algorithms in Python using Pyspark for managing Big dataset parallel processing.

Himanshu’s AI Stack

antigravity
antigravity
cursor
cursor
claudecode
claudecode

Education

Indian Institute of Information Technology Bhopal

Bachelor of Technology, Information Technology

Indian Institute of Information Technology Bhopal

20184 yrs2022

Join Himanshu and 10k+ other professionals on

Tal Boss

Tal is the app where bosses meet engineers & builders directly, without middlemen

Download Tal Boss

Himanshu Panwar on Tal

Headline

Machine Learning | Ex Observe.AI | AI Agent | Speech AI | Building ASR | Conversational AI

About

Himanshu is passionate for Al, Machine Learning, Deep Learning and can build and deploy machine learning and deep learning models using frameworks like PyTorch, TensorFlow, Keras etc. I am also well versed in programming languages like C++ and Python. An ambitious person with excellent research, time management, and problem-solving skills. I loved to solve problems using machine learning which is very helpful and enhance the lifestyle of peoples in the real world.

Location

Bengaluru

AI stack

antigravity

cursor

claudecode

Work experience

Senior AI Engineer at VerbaFlo.AI, Dec 2025 - Present · 9 mos, Researched and deployed task-specific Small Language Models (SLMs), improving reliability, accuracy, and latency, built an intent detection pipeline that reduced latency from 18s to 6s with significant cost savings. Integrated multiple LLM providers (Cerebras, Groq, Gemini, Bedrock) and developed an in-house cross-model prompt benchmarking and evaluation framework, resolving 800+ tickets and remediating 300+ prompts flagged by the benchmark system. Architected a centralized prompt repository with configuration management, fallback logic, validation layers, and automated update pipelines, streamlining prompt operations and accelerating customer escalation resolution.

Senior Machine Learning Engineer at Observe.AI, Nov 2024 - Dec 2025 · 1 yr 2 mos, Bengaluru, Built a real-time transcription enhancement Python library for transcript correction, optimized for high-throughput production traffic with ~3ms per-call latency. Led and delivered the Multilingual Self-Serve Transcription Correction feature in the Observe.AI core product, enabling users to correct transcripts directly from the dashboard across all supported languages, significantly reducing transcription-related escalations. Implemented LLM-based transcription enhancement engine to generate correction suggestions at scale for bulk data, improving transcript quality and reducing manual effort for large datasets.

Machine Learning Engineer at Observe.AI, Jun 2022 - Oct 2024 · 2 yrs 5 mos, Banglore, Built and productionized large-scale ASR systems, iteratively improving in-house models through multi-domain data ingestion, evaluation pipelines, and data cleaning to boost accuracy. Trained large-scale deep learning models using distributed training across A100, V100, and T4 GPUs on AWS and Oracle Cloud using PyTorch DDP, Horovod, and Slurm. Optimized models and inference pipelines using ONNX, Triton, TensorRT, and ESPNET_ONNX, achieving ~75% memory reduction and improved latency in production. Revamped speaker diarization for contact-center mono calls, enabling agent–customer separation, customer-specific customization, and R&D for multilingual end-to-end diarization. Developed an internal speech evaluation toolkit supporting ASR and diarization accuracy evaluation, transcription parsing, normalization, multi-language support, and other functionalities. Developed a customer onboarding and support automation platform deployed across multiple clusters, providing customer-specific analysis, recommendations, enhancements, and evaluation capabilities. Owned and maintained multiple speech services, handling production reliability, monitoring, Python upgrades, and on-call incident resolution with customer-facing impact.

Machine Learning Intern at Observe.AI, Sep 2021 - Jun 2022 · 10 mos, Led R&D of end-to-end ASR systems using ESPnet (Transformer, Conformer, RNN-T, Conformer-Transducer), classical models in Kaldi and self-supervised learning techniques; Successfully trained world-class acoustic models for ASR using large-scale in-house data for contact center. Trained large-scale E2E-ASR models for production on SageMaker, EC2 Ubuntu instances, and Oracle Cloud. Generated over 100K+ hours of speech training data for ASR, including both golden and silver labeled datasets. Developed domain-specific golden labeled test data to rigorously assess the robustness of our in-house ASR system. Designed and deployed Voice Activity Detection (VAD), reducing audio processing by ~40%, doubling throughput, and significantly lowering production cost. Developed and trained models using the traditional speech recognition system Kaldi. I have hands-on experience with ESPnet, an open-source ASR toolkit.

Machine Learning Intern at Mando, Jun 2021 - Aug 2021 · 3 mos, C++ code optimization. Code profiling.

Training and Placement Cell Member at training and placement cell, iiit bhopal, May 2021 - May 2022 · 1 yr 1 mo

Machine Learning Intern at Maulana Azad National Institute of Technology, May 2021 - Jul 2021 · 3 mos, Implemented Cluster-based approach for finding the correlation between the dataset.  Performed multiple experiment i.e. distance correlation, differential privacy algorithms, MIC and calculated information loss.  Developed Algorithms in Python using Pyspark for managing Big dataset parallel processing.

Education

Bachelor of Technology, Information Technology, Indian Institute of Information Technology Bhopal