AI & Model Operations
Every agent in the BFIU platform, the model behind it, and what it is permitted to do · all inference runs on infrastructure inside Bangladesh Bank
Agents deployed
11
across the lifecycle
Models served
8
all on-premises
Inference, 24h
41.2k
calls
Analyst acceptance
78%
of advisory output
Agents with drift flag
2
pending review
Autonomous decisions
0
by design
Stage
Narrative classifier
AGT-01
Reads the narrative an organisation lodged and proposes typology labels, indicator tags and a suspicion-quality score. The analyst accepts or rejects each label individually.
ADVISORYIntake
Model
bfiu-narr-cls v4
Acceptance
81%
Entity resolution
AGT-02
Proposes that two subjects across different reports are the same person or entity, using identity documents, date of birth, address and account overlap.
ADVISORYIntake
Model
bfiu-entres v3
Confirmed
78%
Bangla narrative translation
AGT-03
Renders Bangla narratives into English for analysts working in English, keeping the Bangla text as the record of what was lodged.
ADVISORYIntake
Model
bfiu-bn-en v2
BLEU
41.2
Duplicate and near-duplicate detection
AGT-04
Flags reports that describe activity already reported, whether by the same organisation or a different one.
ADVISORYIntake
Model
bfiu-dedup v2
Precision
94%
Queue ranking assist
AGT-05
Orders the triage queue by likely investigative value, learned from the dispositions analysts recorded on past reports.
ADVISORYTriage
Model
bfiu-rank v3
Override
9%
Network expansion suggester
AGT-06
Proposes which nodes to expand next on the network canvas, based on where relationships are likely to be informative.
ADVISORYInvestigation
Model
bfiu-graph v2
Adoption
63%
Open-source research
AGT-07
Searches indexed open sources for material about a subject, returning each item with its source, date and an unverified marking.
ADVISORYInvestigation
Model
bfiu-osint v3
Verified
44%
Behavioural anomaly model
AGT-08
Scores a subject against its own history and a peer cohort, contributing a named factor to the composite score.
SCORES ONLYDetection
Model
bfiu-anom v2
Contribution
+14 max
Intelligence report drafter
AGT-09
Drafts intelligence product sections from the case record, citing only verified material, for the analyst to rewrite.
DRAFTS ONLYProduct
Model
bfiu-draft v2
Retained
58%
Cluster proposal
AGT-10
Groups subjects, accounts and reports into candidate clusters and states the basis for each proposed link.
ADVISORYInvestigation
Model
bfiu-cluster v2
Confirmed
52%
Supervisory quality reviewer
AGT-11
Flags findings that appear unsupported by cited evidence, for the supervisor to check before review.
ADVISORYReview
Model
bfiu-qual v1
Agreement
61%
Autonomy ladder
ADVISORY9 agents
Proposes something a human accepts or rejects. No effect on any object until a person acts.
SCORES ONLY1 agent
Contributes a named factor to a composite score. Cannot raise an alert or change a state.
DRAFTS ONLY1 agent
Produces text a human rewrites and signs. Cannot mark complete or submit.
AUTONOMOUS0 agents
Would act without a human. Not implemented, and not on the roadmap.
No agent in this platform holds a disposition power. Closing an alert, writing a finding and approving a product are human acts by design, not by omission.
Serving infrastructure
Location
Bangladesh Bank DC, Dhaka
Serving runtime
vLLM 0.6 · Triton 24.08
GPU
8 × A100 80GB
Network
Air-gapped from the internet
External API calls
None
Failover
Warm standby, Rajshahi DR site
Model storage
Encrypted at rest, hash-verified
Inference logging
Full · 10-year retention
Data boundary
No report narrative, subject identifier or transaction record leaves Bangladesh Bank infrastructure. There is no external API call in any inference path. Where a model was trained elsewhere, the weights were brought in and are served locally — the data never travels the other way.