Full-Stack Applied AI Engineer0
Aviral SharmaAviral SharmaApplied AI Engineer
( Bangalore, India — 2026 )

Full-StackAppliedAI Engineer

I design end-to-end AI systems — from autonomous agents and computer vision models to the MLOps pipelines that keep them reliable in production.

Scroll to explore
Agentic AIComputer VisionMLOpsLLM OrchestrationRAG SystemsProduction MLAgentic AIComputer VisionMLOpsLLM OrchestrationRAG SystemsProduction ML

Selected Work

01

Accessibility Audit Agent

Autonomous WCAG 2.2 Auditor

Confidential · OnFinance AI
2026
02

Metering & Entitlements

LLM FinOps Platform

Confidential · OnFinance AI
2026
03

Reporting & MIS

Regulatory Disclosure Workflow

Confidential · OnFinance AI
2026
04

Platform & GitOps Delivery

ComplianceOS Cloud Infrastructure

Confidential · OnFinance AI
2026
05

AI Email Agent

Autonomous Outreach System

2025
06

Trading Agent

Autonomous Financial Analysis

2025
07

AI Research Agent

Semantic Document Intelligence

2025
08

OneARVO Vision

Anti-Counterfeiting & MLOps Platform

Confidential · OneARVO Ventures
2024
AI Email Agent:GitHubLive Demo
Trading Agent:GitHubLive Demo
AI Research Agent:GitHubLive Demo
Aviral Sharma
Aviral Sharma

( About )

I build autonomous AI agents  and the production systems that keep them reliable, governed and scalable.

I'm a Full-Stack Applied AI Engineer. I design end-to-end systems that pair LLM-driven agents and computer-vision models with the backend that makes them production-grade — orchestration, usage metering, observability and MLOps.

Recently I've built multimodal agentic pipelines on LangGraph (deterministic checks fused with LLM vision), an LLM token-metering and entitlements platform for cost governance, and retrieval-augmented document systems — owning the full lifecycle from research and modelling to real-time deployment, resilience and monitoring.

97.9%
CV verification accuracy
56
Compliance criteria auto-audited
65%
Inference latency reduced

( Education )

B.Tech, Information Technology

ABES Institute of Technology (ABESIT) — 2024

Focused on deep learning and computer-vision research, and served as Treasurer of the Bit-Brain Club.

( Toolkit )

PythonGoTypeScriptPyTorchLangGraphLangChainLLM VisionRAGLiteLLMOpenMeterRedisPostgresMongoDBQdrantFAISSAWS (S3 · SageMaker · Bedrock)DockerKubernetes (EKS)ArgoCDKustomizeRabbitMQGitHub ActionsMLflowLangfuseNext.js · ReactVision Transformers

Experience

Applied AI Engineer

OnFinance AI

Generative-AI compliance platform (ComplianceOS) for the BFSI sector.

Bengaluru, IN

Feb 2026 — Present

  • Architected an autonomous WCAG 2.1/2.2 (Level A & AA) accessibility-audit agent — a 14-node LangGraph pipeline that audits Android apps across 56 success criteria by fusing deterministic pixel math (contrast, target-size), accessibility event-log parsing, and multimodal LLM vision (screenshot + UI hierarchy).
  • Scaled that agent to production: adaptive async concurrency (10→25 workers by load), a circuit breaker, per-node checkpoint/resume, and a fire-and-forget S3 watcher daemon (Docker / Kubernetes) that auto-dispatches one audit per capture session.
  • Cut false positives with confidence gating, two-pass LLM verification, and hierarchy pre-filters that remove ~60% of wasted vision calls; cross-screen dedup collapses repeated defects into one canonical finding, with a per-session LLM cost ledger for FinOps.
  • Built an LLM usage-metering & entitlements platform — a stateless Go 1.23 microservice bridging the compliance agents to OpenMeter for per-customer / per-agent / per-model token metering, quotas, RPM/TPM rate limits and billing-readiness, backed by Postgres (pgx).
  • Engineered the metering platform for resilience: fail-open admission (a metering outage never blocks an agent run), a durable Redis outbox with at-least-once delivery and idempotent dedup, and Postgres advisory locks serializing usage across replicas — surfaced through a React + Vite + Recharts admin console behind Dex/OIDC.
  • Delivered the recurring regulatory Reporting & Disclosure (MIS) module — table + calendar tracking of periodic filings with due dates, frequency, owners and SPOC, scoped per business unit and regulator and traced to source circulars, backed by MongoDB + Qdrant semantic search.
  • Owned GitOps delivery for ComplianceOS on EKSKustomize overlays reconciled by ArgoCD across staging, UAT and prod; exposed and OIDC-secured the metering console via ALB + ACM TLS + Dex, with secrets pulled at runtime from AWS Secrets Manager (External Secrets) and Namecheap DNS pointed at the shared ALB.
  • Hardened and operated platform services: environment-locked metering so shared OpenMeter + Postgres never cross-contaminate across envs, removed the admin API from the edge (proxied in-cluster), and shipped a RabbitMQ checklist pipeline (poller → agent → Bedrock via LiteLLM) plus an investigation views-refresh service with health probes, spot scheduling and read-only root filesystems.

Machine Learning Engineer

OneARVO Ventures

Noida, IN

Dec 2024 — Feb 2026

  • Built Vision Transformer and contrastive-learning models on 200K+ images for QR and copy-detection-pattern verification, reaching 97.9% accuracy.
  • Designed automated data curation and perception pipelines to detect, crop and augment QR regions for large-scale training on AWS SageMaker.
  • Architected end-to-end ML pipelines with distributed training, hyperparameter tuning and experiment tracking via SageMaker, MLflow and Airflow.
  • Built real-time inference with model versioning and monitoring, cutting average prediction latency by 65%.
  • Implemented MLOps and observability — CI/CD, drift monitoring and feedback loops with GitHub Actions, CloudWatch and Grafana.

Flutter Developer

BlueTrans

Noida, IN

May 2024 — Nov 2024

  • Integrated the Razorpay payment gateway into an Android app for secure transactions, improving successful payments by up to 70%.
  • Designed app UI to enhance user experience and engagement.
  • Implemented authentication, improving security and reducing unauthorised access.
  • Increased app speed by 30% for a smoother experience across 400 daily active users.

( Contact )

Let's buildsomething meaningful.