Mohit Sharma profile photo 1 on Tal

Mohit Sharma

Hallucinating with agents

Mumbai Metropolitan Region

9+ years of work experience

AjnaLens
Jul 20251 yr 4 mosOct 2026

Thane, Maharashtra, India

Owned the backend and AI platform for production smart glasses, including multimodal agents, device services, and OTA updates. • Led a seven-person backend and platform team delivering the AI and backend architecture for smart glasses serving 600 daily active users and ~4,800 voice sessions per day. • Architected a deterministic state-management layer shared across eight production voice features; improved tool-call adherence by 52% and reduced system-instruction tokens by 40% by moving workflow state out of prompts. • Reduced first-turn TTFB from 3 seconds to under 1 second with pooled Gemini WebSocket sessions, and per-turn TTFB to ~600 ms through asynchronous execution and intent-triggered tool prewarming. • Shipped voice translation, image-based nutrition logging, health-data retrieval through Apple HealthKit and Google Health Connect, and voice-controlled reminders. • Implemented persistent cross-session voice-agent memory using Google's Memory Bank. • Validated 150 concurrent voice sessions per 2-vCPU/4-GB Kubernetes pod; implemented connection-aware KEDA autoscaling, session affinity, connection draining, and graceful shutdown for stateful WebSocket workloads. • Built an automated voice-agent regression and smoke-testing platform covering eight features across English and Hindi, reducing end-to-end testing from 4 hours across three QA engineers to an 8-minute automated job (30× faster feedback). • Identified unintended tool calls under noisy audio through automated testing and introduced a trained RNNoise-based neural noise suppressor. • Evaluated 100% of production sessions with a Langfuse LLM judge; increased agreement with QA labels from Cohen's κ 0.36 to 0.70. • Implemented KMS-backed application-level encryption for personal and conversation data, with decryption restricted to authorized request paths. • Built APIs for authentication, user and device management, OTA updates, and event-driven notifications using Pub/Sub.
Oct 202410 mosJul 2025

New Delhi

Built an agentic learning platform used by approximately 200 users that organizes ChatGPT searches into memorable notes and supports learning through mnemonics and spaced repetition. The product remains active and maintained. • Built LangChain-based LLM-agent workflows instrumented with Langfuse for evaluation, observability, and performance tracking. • Designed a multi-agent RAG architecture using multi-shot prompting to generate questions across configurable difficulty levels. • Developed a FastAPI backend with Celery for asynchronous workloads and scheduled jobs. • Built the frontend with SvelteKit, Tailwind CSS, and shadcn components. • Automated testing and delivery with GitHub Actions. • Deployed the platform with Docker Swarm and Terraform-managed infrastructure.
Hunch
Mar 20248 mosOct 2024

Delhi, India

* Built an ML platform on AWS SageMaker for training, deploying, and monitoring recommendation models. * Built a YAML-driven Airflow framework on AWS Batch for running ML workloads. * Built a production computer vision pipeline for face matching and facial-attribute extraction. * Built CI/CD and Terraform-based infrastructure for ML workloads. * Worked on reducing AWS infrastructure costs and simplifying access management.
Twin Health India
Sep 20212 yrs 7 mosMar 2024

On-site

* Improved production ML model serving, reducing API latency using Redis, model quantization, and asynchronous FastAPI. * Built tooling for parallel ML workloads on AWS Batch with results stored in S3. * Built deployment, monitoring, and alerting for production ML models. * Built shared Python tooling for data access across MySQL, DynamoDB, and Athena. * Built reusable ML project tooling and data pipelines processing 1M+ records
AstraZeneca India
Jul 20174 yrs 3 mosSep 2021

Chennai

* Worked on financial forecasting models using gradient boosting and ensemble methods in R. * Built R Shiny dashboards for sales forecasting and business reporting. * Built an ML model and dashboard for predicting data-center CO₂ emissions. * Built MySQL and Excel ETL pipelines to automate capacity reporting.
GitHub
@mohit2152sharma
Personal portfoliomontepy.in

Mohit’s AI Stack

claude code
claude code
hugging face
hugging face
langchain
langchain
langfuse
langfuse
Pipecat
Pipecat
PydanticAI
PydanticAI

Education

Indian Institute of Technology, Madras

Bachelor of Technology (BTech), Civil Engineering

Indian Institute of Technology, Madras

20134 yrs2017

High School

Vijayvergia Bal Vidyalaya

20112 yrs2013

Join Mohit and 10k+ other professionals on

Tal Boss

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

Download Tal Boss

Mohit Sharma on Tal

Headline

Hallucinating with agents

About

I’m a Machine Learning and AI Platform Engineer with 9+ years of experience building production ML systems, real-time AI agents, and scalable backend platforms. My recent work focuses on taking agentic and multimodal AI systems from architecture to production — including real-time voice agents, LLM evaluation, RAG, persistent agent memory, multimodal workflows, and infrastructure for stateful AI workloads.

Location

Mumbai Metropolitan Region

AI stack

claude code

hugging face

langchain

langfuse

Pipecat

PydanticAI

Work experience

AI/Backend Lead at AjnaLens, Jul 2025 - Oct 2026 · 1 yr 4 mos, Thane, Maharashtra, India, Owned the backend and AI platform for production smart glasses, including multimodal agents, device services, and OTA updates. • Led a seven-person backend and platform team delivering the AI and backend architecture for smart glasses serving 600 daily active users and ~4,800 voice sessions per day. • Architected a deterministic state-management layer shared across eight production voice features; improved tool-call adherence by 52% and reduced system-instruction tokens by 40% by moving workflow state out of prompts. • Reduced first-turn TTFB from 3 seconds to under 1 second with pooled Gemini WebSocket sessions, and per-turn TTFB to ~600 ms through asynchronous execution and intent-triggered tool prewarming. • Shipped voice translation, image-based nutrition logging, health-data retrieval through Apple HealthKit and Google Health Connect, and voice-controlled reminders. • Implemented persistent cross-session voice-agent memory using Google's Memory Bank. • Validated 150 concurrent voice sessions per 2-vCPU/4-GB Kubernetes pod; implemented connection-aware KEDA autoscaling, session affinity, connection draining, and graceful shutdown for stateful WebSocket workloads. • Built an automated voice-agent regression and smoke-testing platform covering eight features across English and Hindi, reducing end-to-end testing from 4 hours across three QA engineers to an 8-minute automated job (30× faster feedback). • Identified unintended tool calls under noisy audio through automated testing and introduced a trained RNNoise-based neural noise suppressor. • Evaluated 100% of production sessions with a Langfuse LLM judge; increased agreement with QA labels from Cohen's κ 0.36 to 0.70. • Implemented KMS-backed application-level encryption for personal and conversation data, with decryption restricted to authorized request paths. • Built APIs for authentication, user and device management, OTA updates, and event-driven notifications using Pub/Sub.

Founder at Saral, Oct 2024 - Jul 2025 · 10 mos, New Delhi, Built an agentic learning platform used by approximately 200 users that organizes ChatGPT searches into memorable notes and supports learning through mnemonics and spaced repetition. The product remains active and maintained. • Built LangChain-based LLM-agent workflows instrumented with Langfuse for evaluation, observability, and performance tracking. • Designed a multi-agent RAG architecture using multi-shot prompting to generate questions across configurable difficulty levels. • Developed a FastAPI backend with Celery for asynchronous workloads and scheduled jobs. • Built the frontend with SvelteKit, Tailwind CSS, and shadcn components. • Automated testing and delivery with GitHub Actions. • Deployed the platform with Docker Swarm and Terraform-managed infrastructure.

Machine Learning Engineer at Hunch, Mar 2024 - Oct 2024 · 8 mos, Delhi, India, * Built an ML platform on AWS SageMaker for training, deploying, and monitoring recommendation models. * Built a YAML-driven Airflow framework on AWS Batch for running ML workloads. * Built a production computer vision pipeline for face matching and facial-attribute extraction. * Built CI/CD and Terraform-based infrastructure for ML workloads. * Worked on reducing AWS infrastructure costs and simplifying access management.

Machine Learning Engineer at Twin Health India, Sep 2021 - Mar 2024 · 2 yrs 7 mos, On-site, * Improved production ML model serving, reducing API latency using Redis, model quantization, and asynchronous FastAPI. * Built tooling for parallel ML workloads on AWS Batch with results stored in S3. * Built deployment, monitoring, and alerting for production ML models. * Built shared Python tooling for data access across MySQL, DynamoDB, and Athena. * Built reusable ML project tooling and data pipelines processing 1M+ records

Data Scientist at AstraZeneca India, Jul 2017 - Sep 2021 · 4 yrs 3 mos, Chennai, * Worked on financial forecasting models using gradient boosting and ensemble methods in R. * Built R Shiny dashboards for sales forecasting and business reporting. * Built an ML model and dashboard for predicting data-center CO₂ emissions. * Built MySQL and Excel ETL pipelines to automate capacity reporting.

Education

Bachelor of Technology (BTech), Civil Engineering, Indian Institute of Technology, Madras

High School, Vijayvergia Bal Vidyalaya

Public links

github: https://github.com/mohit2152sharma

montepy.in: https://montepy.in/