BASE CAMP · 1,400 M FIELD DOSSIER № 07 — EST. 2023 27.7172° N · 85.3240° E · +05:45
Shipping production voice agents and RAG pipelines at Leapfrog Technology — Kathmandu.
Kathmandu, Nepal. Working between the Himalayas & us-east-1.
Open to consults, collabs, and the kind of problem nobody has a Stack Overflow answer for yet.
CAMP I · 2,800 M§ 01 — ABOUT THE SURVEYOR
I bridge the theory with the production-grade.
Machine Learning Engineer with 2.5+ years building Generative AI, RAG architectures, and cloud-native AI for fintech and healthcare. My favourite systems are the ones that think in one modality and speak in another — a phone call that becomes a database row, a stack of PDFs that becomes a conversation.
Right now I'm at Leapfrog Technology, deep in AWS Bedrock, Claude, and Amazon Nova. Before that, Fusemachines and E.K. Solutions, where I cut my teeth on retrieval, prompt engineering, and the unromantic art of making transformers fit inside a budget.
Off-hours I live outside — trekking Langtang, following waterfalls, photographing the mountains that raised me. The thing I believe about engineering: you ship better systems if you remember how small you are.
I like working on problems that don't already have a clean answer. Ambiguity is where the interesting engineering lives. If that sounds like your weekend, let's talk.
2.5+
Since late 202310,000+
Across RAG & GraphRAG pipelines95%
On structured extraction benchmarks
Building AI from the roof of the world.
CAMP II · 4,200 M§ 02 — SURVEY PLATES I–IV
PLATE I — FINTECH / VISAAWS · BEDROCK
A four-stage RAG pipeline on AWS Bedrock + Claude Haiku 4.5, with an ID-mapping citation strategy so every answer points to the exact paragraph in the source document. No more hallucinated references.
pipeline.schema4-stage
PLATE II — FINTECH / LENDINGNOVA PRO · PGVECTOR
A Text-to-SQL assistant powered by Amazon Nova Pro with BM25 + pgvector hybrid search, IVFFlat / HNSW indexing, MMR re-ranking, and AWS Guardrails — so loan officers can ask "which applicants missed their third installment?" and get a correct query back, every time.
nl → sql · previewaccepted
-- "which applicants missed their third installment?" SELECT a.id, a.name, a.email FROM applicants a JOIN installments i ON i.applicant_id = a.id WHERE i.sequence = 3 AND i.paid_at IS NULL AND i.due_at < now() ORDER BY i.due_at DESC LIMIT 50;
PLATE III — HEALTHCARE / VOICENOVA SONIC · TWILIO
A real-time IVR agent that picks up the phone, confirms insurance, and schedules follow-ups — built on Amazon Nova Sonic + Twilio, with a Textract → Nova Pro → OpenEMR pipeline behind it. 60% reduction in scheduling overhead.
call · transcript · 00:23live
callerHi, I need to move my follow-up, my daughter's sick.
agentOf course. Dr. Mehta has openings Thursday at 10 or Friday at 2 — which works?
callerFriday at 2, please.
agentBooked. I'll text the confirmation to the number on file.
PLATE IV — FINTECH / ANALYTICSNEO4J · GRAPHRAG
A knowledge-graph-backed retrieval system over 10,000+ financial documents. Entities and relationships extracted with LLMs, stored in Neo4j, and queried with GraphRAG to answer questions traditional RAG can't — "which guarantors are co-signing loans across three or more portfolios?"
knowledge-graph · subsetneo4j
CAMP III · 6,400 M§ 03 — ASCENT LOG
MAY 2025PRESENT
SEP 2024MAY 2025
JAN 2024JUN 2024
OCT 2023JAN 2024
Curious by nature.
Engineer by craft.
CAMP IV · 7,900 M§ 04 — EQUIPMENT MANIFEST
AI · ML frameworks07
Where the thinking happens.
Languages03
Daily drivers.
Web frameworks03
APIs & edges.
Databases04
Where state lives.
Cloud06
AWS, mostly.
Infrastructure05
Plumbing, done right.
SUMMIT · 8,848 M§ 05 — GET IN TOUCH
Have a problem that's a little too
weird for the average
ML shop?
Email me.