What I build
- LLM-powered applications with schema-constrained outputs, validated with Pydantic or Zod.
- Document AI: extraction from scanned clinical, administrative and financial documents.
- Workflow automation with human-review queues for exceptions.
- AI features inside multi-tenant SaaS, with tenant isolation enforced in PostgreSQL.
- On-premise or hybrid model deployments where data cannot leave your infrastructure.
How a project runs
- Scope: document inventory, data sources and the decisions the app must support.
- Build: schema design, the extraction pipeline, a validation harness and the review interface.
- Harden: accuracy tests on real documents, failure-mode review and audit logging.
- Hand over: code, runbooks and handover documentation.
Proof from production
- iCare: a patient registry and cold-chain inventory platform deployed across state hospital networks in Punjab and Madhya Pradesh.
- TaskEdge: a multi-tenant operations SaaS with AI planning features, in active use.
- Extraction pipelines: over 99.4% field-level precision on invoices and clinical documents.
Stack and constraints
The stack is chosen per project and confirmed during scoping. The common building blocks are Pydantic and Zod validation, PostgreSQL with Row-Level Security for tenant isolation, schema-constrained LLM generation, and local or hybrid model deployment when data privacy requires it.
Engagement formats
Work is sold as a fixed-scope AI integration sprint, a monthly advisory retainer, or a milestone-based build contract. The engagement packages list the deliverables for each. For a readiness review before you build, see AI consulting.