For hiring managers
I'm Joshua Vaz, a Forward Deployed Engineer in Mumbai. I build multi-agent RAG systems and agentic automation for enterprise teams including Coca-Cola. I ship my own products: Kahani Ghar, a kids' stories app live on Google Play with paying subscribers, and Riya, a consumer LLM app that reached 110+ users in five days and is still live. In the last month I added 900 LinkedIn followers, and that number is still growing.
I'm looking for a team that ships production AI, not demos. The work I want is the kind where you sit with the customer, take the constraints seriously, and leave behind a system that holds when real people depend on it.
50,000+
documents indexed in a production RAG pipeline
~99%
cut in information-lookup time versus manual review
12 agents
in one production accounts-payable system
< 2 min
to process an 800-line invoice, with a full audit trail
93%+
cut in healthcare referral handling, ~30 minutes to under 2
4 in 8 weeks
products shipped to production, 350+ users collectively
+900
LinkedIn followers added in the last month, and growing
110+
users on Riya in the first five days
Problem, what I built, outcome. Stacks are on /work.
Forward Deployed Engineer · Feb 2026 to present
Problem
Teams were manually searching more than 50,000 unstructured documents to answer routine questions.
What I built
Architected and deployed a production multi-agent Retrieval-Augmented Generation pipeline over the full corpus.
Outcome
Cut information-lookup time by approximately 99% versus manual review.
Global logistics enterprise
Problem
Accounts-payable invoices were processed by hand, with duplicate-payment risk and no audit trail.
What I built
End-to-end AP pipeline: document-AI invoice ingestion, multi-tier reference resolution, rate-card auditing, and automated posting into an on-premises ERP. Twelve coordinated agents. Idempotency keys throughout.
Outcome
800-line invoices processed in under 2 minutes. Full audit trails. Duplicate-payment risk eliminated.
US healthcare provider
Problem
Referrals required manual work across an EMR, an insurance portal and cloud fax. ~30 minutes each.
What I built
An agentic system spanning all three surfaces, with HIPAA-conscious in-memory PHI handling and a human-in-the-loop exception queue.
Outcome
Per-referral processing cut from ~30 minutes to under 2 minutes, a 93%+ reduction.
A bedtime-stories app for kids, in Hindi and English. I built the Android app, the backend and the content pipeline end to end, and launched it on Google Play.
Paying subscribers · shipped solo
Google Play →A consumer LLM app I built and launched independently. Hierarchical memory: raw chat to per-session summaries to a summary-of-summaries in the system prompt. No vector database, no embeddings.
110+ users in the first 5 days
riya-ai.siteI ship, then I harden. Four products to production in eight weeks taught me the difference between shipping fast and shipping carelessly. The first version goes out. The second version has the idempotency keys.
I work where the client is. Forward deployed means the constraints are real: an on-premises ERP nobody wants to touch, a cloud fax machine in a healthcare workflow, a compliance rule that rules out the obvious architecture. I build around them rather than arguing with them.
I translate. Most of the people paying for AI systems can't evaluate them. I write, in plain English, about what actually goes wrong, and it's made me better at the part of the job that happens in meetings rather than in the editor.
I finish things. I once inherited a live Stripe and Cashfree integration that was dropping payments. I stayed in it until webhook idempotency and transaction integrity were solid. Unglamorous. The kind of work a team actually feels.
Email hello@joshuavaz.com, or find me on LinkedIn. I read both myself.
I also write in public. Read the articles.
