Who I am
About Me
I'm a Senior AI/ML Engineer with 10+ years of experience building production-grade ML systems, Generative AI pipelines, and large-scale analytics platforms. Currently at UST, delivering LLM-powered multi-tenant chatbot platforms, document-aware RAG systems, and Vision AI pipelines for clients like Nokia and AT&T.
My work spans Telecom (2+ yrs) — ReAct/MRKL agentic architectures, Causal Double ML, Graph Neural Networks, and Vision AI for network optimisation — Cybersecurity (2+ yrs) — anomaly detection, ETL pipelines, and AWS-orchestrated analytics at Accenture — and LifeSciences (6 yrs) — genomics Knowledge Graphs, RAG pipelines over genomic literature, biomarker ML models, and clinical trial data engineering at TCS.
I hold an M.Tech in AI/ML from BITS Pilani (A Grade, 2024), with proven impact: reduced Vision AI pipeline runtime by 83%, cut tool integration from 1 week to 1–2 days, and reduced false tool calls by 40% through architectural improvements in production systems.
yrs
Technical Arsenal
Skills & Tools
Academic Background
Education
Career Journey
Work Experience
- Re-architected the core agent from monolithic LangGraph to a ReAct + MRKL-based architecture, enabling parallel and sequential multi-tool execution — decoupling tool orchestration from business logic and improving throughput on complex multi-step queries.
- Cut tool integration time from 1 week to 1–2 days by building a JSON-configuration-driven tool registry — developers define only a config schema and request formatter; all routing, orchestration, and validation handled automatically by the framework.
- Built a global shared cache layer (30-min TTL, cross-session) with pre-dispatch parameter validation — reduced false tool calls by ~40% and lowered downstream server load for better scalability.
- Replaced all hardcoded intent-matching logic with 10-turn rolling conversation context, enabling robust handling of ambiguous multi-turn queries without rule-based fallbacks.
- Architected a production-grade, document-type-aware RAG platform with tailored ingestion strategies per format — Excel/tabular (column, sheet, row granularity converted to header-enriched natural language), PDF/Word (overlapping sliding-window + semantic chunking), and XML (parent-child relationships as chunk metadata).
- Built a multi-technique querying pipeline — Multi-Hop Reasoning, Query Compression, Entity-Based Search, Query Rewriting, and Re-ranking — resolving relational queries via an in-memory knowledge graph built at query time.
- Designed a metadata-driven document lifecycle system with version and document-identity metadata on chunks, enabling targeted deletion/replacement on update and eliminating full re-ingestion.
- Designed a scalable async Vision AI pipeline on GCP to extract telecom pole attributes from 2D/3D street imagery using geo-coordinates, processing thousands of images per batch.
- Reduced pipeline execution time for 4–5K images from ~4 hours to ~40 minutes (~83% reduction) by introducing a Vertex AI Session Pool — distributing inference across concurrent API sessions instead of serialised single-session calls.
- Implemented a Circuit Breaker pattern at the session level — rate-limit failures pause the affected session while traffic reroutes to healthy sessions, preventing cascading failures and full pipeline restarts.
- Designed an end-to-end LLM chatbot on AWS Bedrock for Nokia site engineers — enabling natural-language querying over alarm resolution guides and site status parameters, significantly reducing dependency on customer support.
- Built a secure, encrypted Bedrock Knowledge Base backed by S3 with Amazon Titan embeddings and Amazon Nova Lite for inference, applying semantic chunking for high-fidelity document ingestion.
- Designed a Bedrock Agent with intelligent routing logic to dynamically decide — based on user intent — whether to query the knowledge base or invoke AWS Lambda functions to fetch live site parameters and alarm status in real time.
- Implemented Causal Double Machine Learning (DML) to isolate parameters with true causal impact on spectral efficiency, eliminating spurious correlations from high-dimensional KPI features.
- Addressed class imbalance and noisy labels in telecom fault alarm forecasting, improving F1-score through targeted resampling and label-cleaning strategies.
- Modelled telecom network data as knowledge graphs and trained Graph Neural Network models for anomaly detection and event prediction.
- Architected ETL ingestion pipelines centralising multi-source security telemetry; executed end-to-end migration from MySQL → PostgreSQL → Google BigQuery, improving query performance and analytical scalability.
- Led architecture redesign for real-time sync between PostgreSQL backend and BigQuery frontend using Kafka, maintaining both in sync without downtime.
- Designed statistical anomaly-detection models (isolation forest, z-score baselines, time-series decomposition) and orchestrated pipelines via AWS Step Functions with CloudWatch alerting for model drift and data-quality failures.
- Built EDA and reporting layers translating model outputs into executive-level insight dashboards for security leadership.
- Architected a RAG pipeline over genomic literature using Neo4j as graph store, vector embeddings for semantic retrieval, and an ML synthesis layer — enabling natural-language Q&A over biological datasets; co-authored white paper on semantic knowledge graph modelling for clinical genomics.
- Designed Knowledge Graph schemas for genomic entities (genes, proteins, pathways, phenotypes); trained ML models (random forest, gradient boosting) for biomarker classification; deployed via Flask REST APIs with JWT security.
- Built an Azure ITSM chatbot with multi-step dialog, ServiceNow integration, Image-to-Text AI inference, and Cortana voice channel — deployed across MS Teams, Skype, and Direct Line.
- Built an end-to-end RCA pipeline classifying call failures and voicemail drops from TAS, AMF, and MME logs across millions of call events — reconstructing complete call lifecycles using unique call identifiers and engineering features capturing timing, signalling, and failure-mode patterns.
- Built data sync pipelines for clinical trial digitalization aggregating data from multiple heterogeneous sources with statistical quality checks (distribution tests, outlier detection).
- Implemented GridFS MongoDB streaming API for chunked upload and playback of clinical multimedia data. Led ILP project team — awarded ILP LIREL Best Associate During Training.
Research & Engineering Work
Projects
Genomics Knowledge Graph Platform
End-to-end KG system for LifeSciences genomics research — ingestion, semantic modelling, and interactive visualisation.
AI-Powered ITSM Chatbot
Enterprise chatbot with Image-to-Text AI inference, ServiceNow integration, voice channel, and multi-cloud deployment.
Information Security ETL Dashboard
Cloud-native ETL platform ingesting multi-source security telemetry with AWS Step Functions orchestration and live analytics.
Clinical Trial Digitalization
Multi-source data synchronisation system and GridFS MongoDB streaming API for digitising clinical trial workflows.
LLM-Based Multi-Tenant Chatbot Platform
ReAct + MRKL agentic architecture with JSON-config-driven tool registry, cross-session cache, and rolling context — reducing tool integration from 1 week to 1–2 days and false tool calls by 40%.
RAG Ingestion & Lifecycle Management Platform
Production-grade, document-type-aware RAG platform with tailored ingestion strategies per format (Excel, PDF, XML), multi-technique querying pipeline, and metadata-driven document lifecycle management — eliminating full re-ingestion on updates.
Vision AI — Pole Feature Extraction Pipeline
Scalable async GCP pipeline extracting telecom pole attributes from 2D/3D street imagery. Achieved 83% runtime reduction (4 hrs → 40 min) via Vertex AI Session Pool and Circuit Breaker pattern for fault-tolerant inference at scale.
AWS Bedrock LLM Chatbot — Site Alarm & Status Assistant
End-to-end LLM chatbot on AWS Bedrock for Nokia site engineers — intelligent agent routing between a secure Knowledge Base and live Lambda-fetched site data, eliminating repetitive customer support escalations.
Recognition
Awards & Honours
Get in touch