I build software where the difficulty sits in the system layer: multi-model AI orchestration, event-driven backends, computer vision pipelines, and the business workflows underneath them.

Most of my work is end-to-end product engineering for B2B platforms. That means taking a fragmented manual process, designing the data model and execution architecture around it, integrating AI where it earns its place, and shipping it as a product teams operate daily.

• AI in production with multi-provider LLM orchestration, agent frameworks, MCP tooling, RAG, computer vision, and voice
• Platform architecture across event-driven microservices, durable workflow execution, queueing, multi-tenancy, and RBAC
• Full-stack delivery with Next.js/React surfaces backed by NestJS, FastAPI, and Node services
• Domains spanning growth automation, HR tech, logistics, healthcare operations, ESG reporting, physical security, and e-commerce

AI / ML

Computer Vision

Backend & Distributed Systems

Frontend

Data & Storage

Cloud & Operations

Client engagements, described without naming names.

• Problem: Growth teams ran lead research, outreach, content, CRM, and analytics across fragmented tooling.
• Built: A unifying platform: multi-model AI agent framework with 130+ specialized agents, drag-and-drop workflow builder with 50+ action nodes and approval gates, custom MCP servers for extensible AI tooling, and 30+ third-party integrations.
• Scale: Event-driven microservices sustaining 1,000+ concurrent workflow executions.

Next.js React NestJS Kafka MongoDB ClickHouse LangGraph Kubernetes

• Problem: A passive camera estate recording footage nobody watched, with no active safety or security value.
• Built: Model-development foundation and production inference architecture. YOLO detectors fine-tuned per site for people, PPE, vehicles, plates, fire and smoke. Temporal transformer models for behaviour a single frame cannot resolve. ANPR with perspective rectification and regional syntax validation.
• Scale: Multi-model orchestration running up to 20 models against one video job, deployable to cloud, on-premise, or edge.

Python PyTorch YOLO VideoMAE ONNX TensorRT FastAPI PostgreSQL

• Problem: Sales reps spent 70% of their time dialing and leaving voicemails instead of closing.
• Built: An autonomous outbound calling platform handling introductions, qualification, objection handling, and calendar booking, with a post-call layer for collateral delivery, calendar invites, and sentiment analysis.
• Result: Campaign setup dropped from 2+ hours to under 10 minutes. Live campaigns run at a 25% pickup rate with meetings booked and collateral delivered without a manual call.

React FastAPI MongoDB LangChain ElevenLabs Twilio

• Problem: Merchants ran stores through disconnected tools, manual processes, and specialist contractors.
• Built: A full-stack AI automation platform with nine modules covering support, product page generation, ad creation, social scheduling, and profit analysis.
• Result: Over 90% of routine tickets handled without human intervention. A product URL becomes a review-ready ad campaign in under 15 minutes.

Next.js React TypeScript Supabase Google Gemini n8n

• Problem: Sustainability data lived across spreadsheets, invoices, receipts, and databases, reconciled by hand into reporting templates.
• Built: A platform ingesting unstructured and semi-structured data without forcing fixed templates, using GPT-4 schema mapping with review safeguards, connectors for PostgreSQL/MySQL/SQL Server/AWS RDS, and a KPI engine generating 20+ workforce and emissions metrics.
• Result: 500-row spreadsheets processed end to end in under a minute.

Next.js Python FastAPI PostgreSQL Azure OpenAI Azure Key Vault

• Problem: Managers reviewed patient call quality by manual sampling and subjective judgement.
• Built: A call intelligence system with webhook ingestion, asynchronous recording processing, and PII masking that strips names, phone numbers, and identity numbers before anything reaches an LLM. Analysis scores agent behaviour and surfaces missed conversions and unresolved complaints across three languages.
• Constraint: Privacy-first by design, since the source material is sensitive medical conversation.

Next.js Python FastAPI Celery PostgreSQL Sarvam STT IndicNER

• Problem: Hiring teams worked across disconnected tools for sourcing, screening, interviews, assessments, and scheduling.
• Built: A multi-tenant platform bringing them into one workflow, with an AI recruitment assistant, automated pre-screening and video interviews, live coding assessments, and durable workflow orchestration.

NestJS NX Monorepo React PostgreSQL PgVector Temporal LangChain AWS

I'm interested in problems where the architecture is the hard part: AI systems, automation platforms, and B2B SaaS.