Intro
I'm a Senior Software Engineer at GEP Worldwide, building the Leo Agentic Runtime and the Leo search platform - agentic streaming protocols, durable Human-in-the-Loop subsystems, a pluggable Elasticsearch query platform, and a Logstash CDC fabric spanning 171 pipelines across 28 business domains. On the side I'm building Continua, a Go/Postgres durable execution engine purpose-built for AI agents.
Experience
Senior Software Engineer (Jul 2025 - Present) · Software Engineer (Jul 2023 - Jul 2025) GEP Worldwide-Mumbai, IndiaJul 2023-Current
- Engineered the AG-UI streaming protocol adapter for the Leo Agentic Runtime - a real-time SSE bridge translating LangGraph executions into typed AG-UI events; cut time-to-first-token from ~5s batch to ~500ms streaming.
- Engineered session-scoped MCP client pooling - 100-client LRU pool with a promise-chain mutex preventing overflow and delete-before-close races; measured 2.8× speedup (1250 ms → 450 ms) on a 5-tool workflow.
- Architected the durable Human-in-the-Loop interrupt/resume subsystem - 4 interrupt types via custom LangChain middleware, MongoDB checkpointer; 100% in-flight survival across pod restarts, P95 < 500 ms resume latency.
- Architected the Leo Search Runtime - a pluggable Elasticsearch query platform exposing 7 query types, a recursive expression evaluator over 6 operators, and a strategy-based text pipeline with 8 normalization strategies.
- Built and operated the Leo Storage CDC platform - a Logstash fabric on secured Kafka (SCRAM-SHA-512), 171 pipelines spanning 28 business domains, keeping Elasticsearch in sync with MSSQL/MongoDB sources across Azure and AWS.
- Shipped semantic search via a KNN dense-vector query builder over Elasticsearch backed by a 3-tier embedding cache - collapsed p95 vectorization latency from ~200 ms to <10 ms on cache hits.
- Designed SearchOps, a REST API for Elasticsearch index lifecycle management - orchestrating ~1k bulk sync/reindex jobs/month across 80+ managed indices.
Machine Learning Intern IHub-Data, IIIT Hyderabad-Hyderabad, IndiaJul 2022-Jan 2023
- Built a Flask web app for the Roadex project running 3 YOLO-based CV pipelines (pothole, triple-rider, tree obstruction) on dash-cam video at ~12 FPS - streaming MJPEG frames and persisting run metadata in SQLite.
- Built the triple-rider violation pipeline cascading TensorFlow YOLOv4 (rider/motorcycle) with PyTorch YOLOv5 (helmet) - 0.83 mAP@0.5, 91% helmet accuracy - using IoU-based matching and Deep SORT tracking.
Machine Learning Engineer Omdena × EnergyHub-Remote, USADec 2021-Feb 2022
- Selected among 50 ML engineers from 20 countries for an Omdena resource project sponsored by EnergyHub - optimized EV distributed network service by analyzing EV charging data.
- Designed clustering algorithms organizing unlabelled time-series data into clusters by electricity consumption profile - impacting 20+ business partners of C4Net and EnergyHub.
Research Intern India Meteorological Department (IMD)-Remote, IndiaNov 2021-Jun 2022
- Engineered an LSTM-based prediction model outputting the next-day severity plot of a hailstorm to occur, from meteorological inputs.
- Implemented mathematical functions and statistical reviews backing a peer-reviewed journal submission.
Research Assistant ISRO-NESAC-Shillong, IndiaSept 2021-Jan 2022
- Designed stacked LSTM and LSTM-CNN models for multivariate time-series forecasting of ionospheric parameters across the North East Region of India - RMSE 0.22.
- Built an LSTM-autoencoder anomaly-detection pipeline for earthquake precursors (F1 0.78) and Mini-SOM zone clusters for 1-day forecasts at 53% accuracy.
- Streamlined a Dash-framework web interface automating the full project pipeline for a minimalistic user experience.
Education
Bachelor of Technology in Computer Science & Engineering National Institute of Technology Agartala-Agartala, India2023
Projects
Durable execution engine for AI agents Continua-Sept 2025-Current
- Architected a Go/Postgres durable execution engine with event-sourced replay, byte-exact divergence detection, lease workers, exponential-backoff retries, continue-as-new, and scope+key CAS dedup; p99 round-trip stays under 50 ms.
- Built a React trace debugger with virtualized 1K+ span waterfalls, state diffs, retry-safety classification, and per-span LLM token/cost attribution.
Python · TensorFlow · NLP DeepHandwritingSynthesis-Jul 2022-Current
- Implemented an attention-based encoder-decoder RNN handwriting synthesizer and studied sequence-to-sequence models from current research literature.
Python · TensorFlow · NLP Forex Trading Recommendation System-Mar 2022-May 2022
- Built a CNN-LSTM recommendation system combining finance-domain Twitter sentiment with FOREX time-series signals, reaching an F1 validation score of 0.62.
Python · OpenCV · TensorFlow OpenCV Sudoku Solver-Dec 2020-Dec 2020
- Built a Sudoku solver with image processing, backtracking, and a fine-tuned VGG-16 digit recognizer reaching 99.3% accuracy.
Capabilities
Languages
English - fluent, Hindi - native
Programming Languages
Go, C#, Python, TypeScript, JavaScript, C/C++, SQL
Libraries & Frameworks
.NET, Angular, React.js, TensorFlow, PyTorch, Flask
Service Providers
Github, AWS
Content Platforms
PostgreSQL, MongoDB, Elasticsearch, Kafka
Tools
Git, Docker, Kubernetes
Certifications
- 2022-Computer Vision Nanodegree - Udacity
- 2021-Deep Learning Specialization - Coursera
References
Interests
I'm a Senior Software Engineer who likes building the infrastructure other people's agents and applications run on - durable execution engines, streaming protocol adapters, search platforms, and CDC fabrics. Outside of work, I'm building Continua, a Go/Postgres durable execution engine for AI agents.