How our AI immigration officer was trained

DOSSIAR's AI officer is not a single model. It's a pipeline: a deterministic evaluator per visa category, a retrieval layer over paraphrased IRCC operational guidance, and an LLM that synthesizes the verdict with citations.

We started with the 23 most-requested Canadian visa categories. For each, we wrote a per-category Evaluator — a small TypeScript module that checks hard eligibility rules deterministically. Funds thresholds, age bands, points totals, refusal histories. Rules that an LLM alone would fumble.

Then we layered on retrieval. Our KB contains paraphrased chunks of public IRCC operational manuals, policy directives, and common refusal reasons. Hybrid vector + BM25 retrieval with reciprocal rank fusion surfaces the most relevant passages for each case.

Finally, the LLM is given the deterministic rule findings PLUS the retrieved policy chunks and asked to produce a structured verdict: approve / refuse / RFE, approval likelihood, ranked concerns, concrete recommendations, and citations back to the KB.

We validated the pipeline against 40 golden cases (20 Study Permit, 20 FSW) handcrafted by licensed RCICs. Verdict agreement is the north-star metric we report every week.