A recruitment platform for Cambodia — semantic matching between candidates and companies, a full hiring pipeline, chat, interviews, and an AI resume builder — on seven NestJS services.
01 / The problem
Apsara Talent has to support two different conversations: candidates and companies finding each other, and a company moving an applicant from application to offer. It needs both without blurring them — and it has to match people on meaning rather than exact keywords, because “Full Stack Development” and “Backend Developer” describe overlapping people.
02 / Key decisions
Decision 01
Recommendations combine weighted signals — career-scope similarity, skill overlap, and job-title similarity — with the semantic parts scored by pgvector cosine distance over an HNSW index, so related roles match even when they share no words.
Trade-off
Not every profile has an embedding yet, so each semantic signal falls back to exact overlap — results stay sensible, just less forgiving, until the vector exists.
Decision 02
A gateway and six internal services — auth, users, jobs, chat, resumes, and notifications — talk over NestJS TCP transport and share one PostgreSQL schema and one contracts library.
Trade-off
The deployables are separate but the data is not: a change to a shared entity can touch several services at once, and the project map says so plainly rather than pretending otherwise.
Decision 03
Every push runs a load phase against the readiness endpoint, which fans out to the database, Redis, and all six services. The p95 limit is 500 ms — about three times the worst of three measured runs.
Trade-off
Concurrency was cut from 20 to 5 after runner contention failed a release at 3.3% errors. A flaky gate gets rerun until it passes and stops meaning anything.
Decision 04
Restoring the database from a Neon point-in-time branch is tested weekly and takes 3.7 seconds; rolling a single service back takes about two minutes, with a dry-run rollback planner in CI.
Trade-off
The point-in-time window is one day, so anything noticed later depends on the nightly off-site dump — written down in the runbook rather than discovered during an incident.
03 / Outcome
Deployed across Railway, Neon, Cloudflare R2, and Vercel, with Prometheus, Alertmanager, and Grafana watching it and a runbook for the failures that have already happened once. Roughly 1,700 commits across the API, web, and mobile repositories since January 2025.
04 / Inside the product
05 / Technologies
06 / Public resources