BLOG
Insights from
Our Team
Engineering deep-dives, research notes, and practical guides from the people building Synapse.
Most of what we write comes out of building Synapse Intelligence, a governed document layer for banks, insurers, and public institutions that cannot move their records to someone else's cloud. OCR over scanned Nepali and English paperwork, retrieval and named-entity recognition in a low-resource language, search engines light enough to live inside the database instead of a separate cluster, multi-tenant access control across branches and head office, deduplication and workflow routing, and answers that stay traceable to the source document. Written by the engineers who did the work, not a content team.
Real time fine-tuning of recommendation systems
Recommenders retrain on a schedule, so they stay blind to whatever changed this week. We built a dataset and used smaller language models to steer ranking in real time, so moderation decisions and stated user preferences take effect without waiting for a full retrain.
By Pranjal Timsina
Scaling Securely - Why Multi-Tenant RBAC is No Longer Optional for SaaS
Tenant isolation and per-tenant roles are the part of a SaaS product that quietly turns into spaghetti. What multi-tenant RBAC has to guarantee before you onboard the next customer, and why we packaged ours as Synapse RealmGuard instead of rebuilding it per product.
By Riyesh Duwal Shrestha
Finetuning NepBerta for Nepali NER
Nepali is a low-resource language, so general-purpose entity models have very little to learn from. We annotated our own dataset and fine-tuned NepBERTa to tag people, organizations, and locations in Nepali text, the groundwork for pulling structured fields out of Nepali documents. The model is public on Hugging Face.
By Priyanshu Koirala
From ElasticSearch to PostgreSQL Vector Search: Optimizing Nepali Language Search
We designed for ElasticSearch and shipped on PostgreSQL. How pgvector, tsvector ranking, and an ILIKE fallback covered Nepali-language search inside the database we already ran, and what keeping the index in one place saved us in operational overhead.
By Riyesh Duwal Shrestha
