SELECTED WORK

Built. Shipped.
Still running.

Production AI systems for enterprise teams, including a Fortune 500 company and one of India's largest conglomerates, alongside consumer products shipped solo.

Coca-Cola · Fortune 500
Adani · Indian conglomerate
Biocair · Pharma cold-chain logistics
Iris Healthcare · US healthcare
Alamance · Document processing

Retrieval engines

Systems that answer questions from a body of documents, instead of making the answer up. The hard part is retrieving the right passage, not generating the sentence.

Coca-Cola

Production RAG engine over 50,000 documents

Python
FastAPI
LangChain
OpenAI API
Claude API
Pinecone
PostgreSQL
Docker

Problem: Teams were searching more than 50,000 unstructured documents by hand to answer routine questions.
Built: A production multi-agent retrieval-augmented generation pipeline over the full corpus, architected and deployed end to end.
Outcome: Information lookup time cut by roughly 99% against manual review.
50,000+ documents indexed and queryable.

Document management software vendor

Search and indexing for an enterprise DMS

Apache Tika
Apache Solr
OCR
Docker

Problem: Search across a large document management system was slow and returned unreliable results.
Built: Backend search and indexing for the GLOBODOX document management system, with a containerised Tika, OCR and Solr stack.
Outcome: Search performance doubled.
Roughly 20% fewer query errors.
Deployment time cut by about 60%, indexing accuracy up about 25%.

Multi-agent pipelines

Several specialised agents coordinating on one job, each owning a step. Useful when a task has stages that need different tools, and a single prompt would be too brittle.

Global logistics enterprise

Twelve-agent accounts payable automation

Python
FastAPI
LangChain
Document AI
PostgreSQL
On-premises ERP integration

Problem: Invoices were processed by hand, carrying duplicate-payment risk and leaving no audit trail.
Built: An end-to-end accounts payable pipeline: document-AI invoice ingestion, multi-tier reference resolution, rate-card auditing and automated posting into an on-premises ERP. Twelve coordinated agents, with idempotency keys throughout so a retry can never pay twice.
Outcome: 800-line invoices processed in under two minutes.
Full audit trail on every posting.
Duplicate-payment risk eliminated.

Enterprise data ingestion

Fuzzy-matching ingestion agents

Python
Fuzzy matching
REST APIs

Problem: Incoming records did not match existing ones cleanly, so history was lost on ingest.
Built: Ingestion agents using multi-tier fuzzy matching to resolve records against existing history.
Outcome: 100% historical recall on matched records.

Human in the loop

Automation that knows what it should not decide alone. A person is routed the cases the system is not confident about, and everything the system did decide can be audited afterwards.

US healthcare provider

Referral orchestration with an exception queue

Python
FastAPI
LangChain
EMR integration
Insurance portal integration
Cloud fax integration

Problem: Each referral took around 30 minutes of manual work across an electronic medical record, an insurance portal and a cloud fax service.
Built: An agentic system spanning all three surfaces, with HIPAA-conscious in-memory handling of patient data and a human-in-the-loop exception queue for cases the system should not decide on its own.
Outcome: Referral processing cut from about 30 minutes to under two minutes, a reduction of more than 93%.
Clinical exceptions routed to a person rather than guessed at.

Document processing

Turning paperwork into structured data: invoices, forms and scans that arrive as images and have to come out as records a system can act on.

Alamance

Intelligent document processing workflow

Document AI
Python
REST APIs

Problem: A recurring document workflow was consuming more than ten hours a week of manual effort.
Built: An intelligent document processing workflow to handle the intake and routing automatically.
Outcome: More than ten hours a week of manual work reduced to under five minutes.

Enterprise integration

The unglamorous layer that decides whether any of the above survives contact with a real company: the APIs, the on-premises systems and the systems nobody wants to touch.

Enterprise deployments

Secure APIs between LLM engines and enterprise dashboards

FastAPI
REST APIs
Authentication

Problem: Model output needed to reach existing enterprise dashboards without opening a new security hole.
Built: Secure REST APIs connecting LLM engines to enterprise dashboards.
Outcome: Model output delivered into systems teams already used.