Services
Four ways to work with Sabouri.org — whether you're a consulting firm building an AI practice or an accounting / professional-services firm bringing AI in‑house over your own books. From a two-week assessment to a full build your own engineers own at the end. Everything runs fully on-premises — for client confidentiality, professional privilege and data-residency, that isn't a preference, it's the entry requirement.
The method
The habit worth installing in a team is not a technology choice — it's how a fix gets closed and how a figure earns the right to appear on a board deck. Ninety days, in the order that keeps the later decisions honest.
Weeks 1–3 · Evidence
One real workflow with a real owner, and — before any model work — a corpus of questions with known correct answers, including the ones the system should refuse. Without it, every later decision becomes taste.
Weeks 4–8 · The spine
Hybrid retrieval with a measured relevance floor, a mandatory visibility filter on every query, and a fail-closed permission model whose failures say which failure occurred. “Found nothing”, “denied” and “nothing ingested” must never render as the same answer.
Weeks 9–12 · Trust
Anything presented as a figure computed in code, not by the model. Citations verified programmatically. Extraction gated by reconciliation against the source document's own totals. Then ship to a real user group and watch what they actually ask.
Ingest
PDF · XLSX · DOCX · EML · OCR → reconcile vs. source totals
Index
Qdrant bge-m3 1024-d + lexical · per-client tag
Visibility filter
mandatory on every query · fail-closed
Hybrid retrieval
rerank + MMR · measured relevance floor
Agent tool-loop
model picks tools mid-turn · bounded preconditions
Deterministic compute
pandas over Postgres · never the model
Citation verify
reachable · in-range · fresh — in code
Verified answer
or a visible refusal — never a confident guess
Representative request path. Anonymized — no client data.
Engagements
Your intended AI architecture reviewed against the failure modes that end engagements, with a written finding and a remediation order. Useful whether or not we build anything.
The ninety-day method above, delivered with your engineers in the code — not watching. You keep the system, the tests, and the reasoning behind every threshold in it.
Design authority and code review across your client engagements, so the second build is cheaper than the first and the fifth is a product.
Everything we build runs with no external API dependency. For regulated, privileged or data-resident work, that is the baseline — not an upgrade.
Support & continued assurance
We don't hand you a system and a phone number. Each build includes an embedded self-coder — a virtual assistant that lives inside the stack, answers questions about how it works, and drafts and applies routine fixes in place, under your approval. It absorbs the day-to-day, so Sabouri.org (your Tier 3) isn't on the critical path for every question — you call us for the things that actually warrant it.
A built-in assistant that explains the system, answers “why did it do that?”, and drafts and applies small fixes against the real codebase — reviewed before anything lands. Your team self-serves instead of waiting on us.
For what the assistant shouldn't do alone — architecture changes, incidents, model and retrieval tuning, security review. You escalate the real things; we don't meter the trivial ones.
Scheduled assurance reviews — the evaluation corpus re-run, thresholds audited, security and dependencies checked — so a system that was correct at hand-over stays correct as the data, the models and the questions change.
Candour
We don't train foundation models. We make existing ones safe to put in front of someone's books.
We write the code, run it in production, and get paged when it's wrong.
No agent that states figures it can't show its work for. If that's the deliverable, we'll say so in week one, not month six.
There's no licence here. You end up owning what gets built.