// 01 · About

10+ years in AI and ML.

// 02 · Experience

Where I've worked.

1. 01Apr 2025 – PresentHyderabad

   Deccan AI

   RL Lead · Reinforcement learning & ML engineering

   • Leadership & roadmap. Part of Deccan AI's leadership team and owner of the yearly RL roadmap. Scaled the AI/RL engineering team from 1 to 10 engineers and run delivery teams across several projects, about 20 people in all.
   • Stark. Simulated Environments, Trajectories and Agentic RL Kit. Led the team that built the RL Gym Studio platform and set its research roadmap for multi-app RL gyms. Hands-on, built the MCP build framework that automatically converts 80%+ of API docs into mock connectors. The kit covers human-in-the-loop annotation for golden trajectories, synthetic data and task generation, a training wrapper for agentic RL runs and trajectory analysis for grading agents. Delivered 100+ environments, averaging 30 tasks each, to a top-five AI lab. Blog ↗
   • Coding RL Gym. 500+ verifiable software-engineering tasks mined from real private repositories, across dataset families such as DeepSWE and Terminal-Bench, for training and evaluating coding agents. Built the repo screening and PII-redaction pipelines, and Manthan, a task-creation pipeline that mines, authors and repairs tasks and automatically creates model-breaking DeepSWE tasks at a 70% automated yield, plus an IDE-driven human-in-the-loop fix flow for the rest. Frontier models such as GPT-6.1 Sol pass fewer than 40% of them.
   • Caption Bench. Led the team that designed the research framework for evaluating video understanding across 7 properties, published as Deccan AI's video-captioning benchmark. The best of the top five VLMs reaches only about 57% accuracy. Research ↗
   • Media Studio. Tech architect and sole backend lead for an OpenRouter-style LLM gateway: many model providers behind one API with routing, rate limiting and failover, serving about 2M requests a day and bursts above 10,000 requests per second.
   • Expert Talent Search. Hybrid retrieval over 2M+ profiles to source domain experts for RLHF projects. Dense embeddings, structured filters and LLM query understanding and reranking reach nDCG of about 0.70. A Celery pipeline parses 1,000+ resumes per batch into the index.
   • AI Interviewer at scale. End-to-end voice-agent interviewer with role-specific questions and adaptive follow-ups. A proctoring model (voice classification, gaze detection and tracking) flags suspicious sessions so interviews run unattended, and an evaluation engine scores candidates across several dimensions.
   • Browser-use & computer-use gyms. Built the first stage of the browser-use and computer-use gyms, and advise the teams now extending them, along with other RL environment, training and evaluation projects.
   • Research. Researching recipes for synthetic data pipelines for coding benchmarks, including DeepSWE-style tasks, formal verification in coding benchmarks, computer-use gyms and design-to-code (UI/frontend) gyms.

   Agentic RLRL gymsMCPverlRLHFLLM gatewaypgvectorLiveKit

2. 02Nov 2022 – Dec 2024Bengaluru / Abu Dhabi

   Mensa Brands

   ML Engineer · Individual contributor

   • Pricing Central. Dynamic pricing with Prophet demand forecasting and Mixed Integer Programming for portfolio-wide price optimisation, delivering a 5-7% EBITDA uplift.
   • Infinity. Serverless MLOps platform on AWS CDK integrating Lambda, Step Functions, Batch and GitHub Actions for automated deployment, training, monitoring and data ingestion.

   ProphetMIPAWS CDKServerless MLOps

3. 03Mar 2021 – Aug 2022Bengaluru

   Johnson Controls

   Principal ML Engineer

   • Business Logic Module. Re-engineered the layer that processes RabbitMQ events and computes derived metrics, for a 10x faster response time and 15% lower deployment cost.
   • ACVS Engine Revamp. On the AI Architecture team, redesigned object detection on YOLOv8, extended it to multi-class detection and multi-object tracking, and moved tracking to DeepSORT.
   • JCI AI Round Table, India. Key contributor to the AI Leaders Committee, driving applied AI strategy and prioritising solutions that shaped the long-term roadmap in the smart ACVS sector.

   YOLOv8DeepSORTRabbitMQAI architecture

4. 04Jan 2019 – Mar 2021Bengaluru

   Mfine (Novocura)

   Senior ML Engineer

   • Enigma ML Platform. Led the design and build of MLOps infrastructure from the ground up on Metaflow, MongoDB, Docker, Kubernetes and DVC, cutting retraining time 50% and AWS cost 75%, with one-click deployment and end-to-end tracking.
   • Maya AI Assistant. Rebuilt the doctor-assistant tool with Bayesian inference for symptom matching and Siamese networks for differential diagnosis, cutting physician workload 30% with an 80% click-through rate on top-5 predictions.
   • Medical Embeddings. Trained domain-specific Word2Vec, GloVe and Bio-BERT models on proprietary clinical data (SNOMED-CT, ICD-10), improving several in-house clinical NLP tasks.
   • Other contributions. A proprietary explainable-AI library (Grad-CAM, CAM, attention inference), and mentoring interns on an MFCC-based heartbeat-classification system.

   MetaflowKubernetesBio-BERTSiamese netsXAI

5. 05Dec 2016 – Jan 2019Bengaluru

   SigTuple

   ML Engineer IV

   • Shonit blood analyser. Computer-vision WBC extraction with segmentation, adaptive thresholding and morphology, improving accuracy from 67% to 99%. Trained 200+ deep models for WBC classification reaching 96% F1, leading to two successful clinical studies.
   • Drishti fundus / OCT screening. Re-engineered the diabetic retinopathy screening pipeline around a feature-pyramid-pooling CNN that significantly outperformed previous benchmarks.
   • Other contributions. Designed the Dipstick analyser pipeline on ORB descriptors and an SVM, and contributed to the in-house ML platform for segmentation, structured data and a scikit-learn wrapper.

   Computer visionCNNSegmentationClinical studies

// 03 · Advisory

Advising startups.

// 04 · Toolchain

Tools I use.

1. 01

   Environments

   • Agentic RL gyms
   • OpenAPI mock environments
   • Synthetic data & task generation
   • PostgreSQL
   • Redis
   • Celery
2. 02

   Training

   • verl
   • RLHF
   • PyTorch
   • TensorFlow
   • Hugging Face
   • CUDA
3. 03

   Evaluation

   • LLM evaluation
   • Trajectory analysis
   • Graders & verifiers
   • Benchmarks
4. 04

   Retrieval & agents

   • RAG
   • Hybrid search
   • LLM reranking
   • pgvector
   • FAISS
   • LangChain
   • LlamaIndex
   • LiveKit voice agents
5. 05

   Serving & infra

   • vLLM
   • NVIDIA Triton
   • FastAPI
   • Docker
   • Kubernetes
   • AWS: Lambda, Step Functions, SageMaker, Batch, Bedrock
   • GitHub Actions CI/CD

// 05 · Papers & education

Papers and education.