Bespoke engagements — when no template fits

Bespoke AI engineering for novel verticals

Cognilium runs 12 templated industry practices — financial, healthcare, technology, construction, education, energy, hospitality, insurance, logistics, manufacturing, marketing, retail, telecom. When a client's vertical does not fit any of those, we run a structured 2-3 week discovery sprint that produces a sized, milestone-gated engagement plan. We have shipped into govtech, agtech, proptech, legaltech, nonprofit, sports analytics, and civic tech.

2-3 week discovery sprint — regulatory landscape map, platform inventory, data-source catalogue, sized engagement plan with risk register. Fixed fee; artefacts are yours either way.
Per-vertical compliance — FedRAMP Moderate/High and CMMC L2 for GovTech, MLS data-licensing for PropTech, ABA Model Rule 1.6 for LegalTech, state-by-state regulators for any wagering work.
Milestone-gated delivery — four explicit stop-or-continue gates (discovery, integration validation, pre-promotion, 30-day live). No creeping commitment.
Same primitives, novel application — RAG and GraphRAG knowledge graphs, agent supervisors, voice AI, document intelligence, computer vision. The building blocks are battle-tested; the data and policy layers are vertical-specific.
Verticals shipped bespoke

Where we have engineered beyond the template

Public sector & civic
GovTech (FedRAMP)Civic tech (311 + GTFS)Nonprofit / Foundations
Specialised commercial
PropTech (MLS + AVM)AgTech (Sentinel-2)LegalTech (matter intel)Sports analyticsAutomotive dealer techMedia & publishing
Track record
50+
Projects delivered
96%
Client satisfaction
2019
Founded

Six bespoke verticals we have shipped into recently

Each was scoped via a discovery sprint and delivered against milestone gates. The platforms are named because the work is real — not because the platforms are a marketing claim.

GovTech & defense ISVs

FedRAMP-bound, CMMC-aware, IL4/IL5 conscious

  • ▸Reference architectures on AWS GovCloud, Azure Government, GCP Assured Workloads, Microsoft 365 GCC-H — boundary selection done in week one, not after the auth-to-operate audit
  • ▸NIST 800-53 control mapping built into the system as a first-class artefact: SSP modules generated from infrastructure-as-code, not maintained in a parallel Word document
  • ▸CMMC Level 2 control implementation for federal-contractor subs; for classified-adjacent work we scope the boundary at IL4 and flag the limit before any code is written
  • ▸Where Salesforce GovCloud is the system of record, we ingest via the Tooling and Bulk 2.0 APIs and never write back synchronously on the citizen-facing path

PropTech & real-estate platforms

MLS data-licensing first, AVM second

  • ▸MLS ingestion over the RESO Web API with per-board IDX/VOW rule honouring — display restrictions, attribution, opt-out lists enforced in the application layer, not patched in after a board complaint
  • ▸AVM (automated valuation model) using CoreLogic property attributes, MLS history, and parcel-level geospatial joins — calibrated per metro, not nationally averaged
  • ▸Property-management integrations: MRI, Yardi Voyager, AppFolio, RealPage — read paths via vendor REST APIs, write paths queued and idempotent
  • ▸Transaction-readiness scoring (likelihood-to-list, likelihood-to-close) for brokerages with explicit fair-housing review: no protected-class proxies in the feature set, disparate-impact tests before deployment

AgTech & precision agriculture

Satellite + sensor fusion, not just dashboards

  • ▸Sentinel-2 (10m, 5-day revisit) and Landsat-8/9 (30m, 16-day) imagery with cloud-mask + atmospheric correction — NDVI, EVI, NDRE, and SAVI indices computed per field-polygon, not per pixel-grid
  • ▸John Deere Operations Center and Climate FieldView telemetry ingestion: planting maps, as-applied prescriptions, yield monitor passes — joined to imagery on field-GUID and season
  • ▸Prescription-planting and variable-rate maps generated against soil sample grids and historical yield zones; MOA (Mode of Action) rotation logic for crop-protection planning
  • ▸Edge inference at the implement: a quantised model runs on a Raven or Trimble controller when connectivity is intermittent, syncs deltas when a cab radio reaches LTE

LegalTech (beyond Paralegent AI)

Matter intelligence over iManage and NetDocuments

  • ▸Clause-level extraction with confidence scoring against firm-specific playbooks — outputs flow into Westlaw / LexisNexis citation checking and the matter-management system, not a separate review tool
  • ▸iManage and NetDocuments document-OS integration: matter-aware retrieval, ethical-wall enforcement at the index level (Bloomberg Law and Relativity supported for e-discovery scopes)
  • ▸Bates-numbering and redaction pipelines that survive auditor review — every redaction is logged with the rule that triggered it, every Bates-number range is reproducible
  • ▸ABA Model Rule 1.6 (confidentiality) honoured operationally: client data never leaves the firm tenancy, model calls are logged with matter ID and retention-policy aware

Nonprofit & foundation analytics

Donor predictive on Blackbaud and NPSP

  • ▸Donor next-gift modelling on Blackbaud Raiser's Edge NXT, Salesforce NPSP, and GuideStar wealth-screening signals — Cox survival models for lapse risk, not naive last-donation-date heuristics
  • ▸Campaign attribution for multi-channel solicitation (mail, email, peer-to-peer, gala) with uplift modelling — measures incremental gift, not credit-grabbing on donors who would have given anyway
  • ▸Grant pipeline scoring against Foundation Directory Online and 990-PF filings — likelihood of fit by funder, with the regulator-mandated UBIT and self-dealing constraints encoded as features
  • ▸Donor-data hygiene pipeline: NCOA address updates, deceased suppression, household consolidation — every nonprofit's data spine before any model is meaningful

Sports analytics & player tracking

Event ingestion to possession-value models

  • ▸Player-tracking ingestion from Sportradar, Stats Perform, Second Spectrum, and league-specific feeds (NFL Next Gen Stats, NBA SportVU successors) — UTC-aligned to 10Hz
  • ▸Possession-value, expected-points-added, and pitch-control models — published values mirror PFF and StatsBomb methodology so analysts can benchmark internally
  • ▸Injury-risk and load-management models on GPS + heart-rate + sleep data — outputs go to S&C staff, never to roster-decision automation (player welfare is a human call, always)
  • ▸Officiating-aware models: penalty rates, foul incidence and bias detection at the crew level — feeds league office, not in-game decisioning

How we engineer a vertical we have not templated

Novel verticals fail when discovery is shallow, regulation is treated as a documentation exercise, and integrations are not validated until the build is half-done. We invert all three.

Discovery sprint — 2 to 3 weeks, fixed fee

How we de-risk a vertical before any production code is written.

  • ▸Week 1: 4-6 subject-matter expert interviews on the client side; regulatory landscape research over the actual statutes and guidance (not summaries); platform inventory of every system the workflow touches
  • ▸Week 2: data-source catalogue with licensing review per source — critical for verticals like PropTech (MLS redistribution rules) and AgTech (sensor-vendor TOS); candidate architecture sketches; build-vs-buy decision on each component
  • ▸Optional week 3: a thin proof-of-concept on the riskiest unknown — usually the data quality or the regulator-facing artefact, rarely the model itself
  • ▸Output is yours regardless of whether you continue with us: the artefacts are deliverables, not engagement bait

Regulatory landscape mapping

Encoded as policy, not as a footnote in the SOW.

  • ▸GovTech: NIST 800-53 control families scored by applicability to the workflow, FedRAMP boundary selected (Moderate vs High) early, CMMC L2 mapped where federal-contractor scope applies
  • ▸PropTech: MLS board-by-board licensing (IDX, VOW, RETS-deprecation, RESO Web API), Fair Housing Act feature-set review, state-level real-estate brokerage rules
  • ▸LegalTech: ABA Model Rule 1.6 (confidentiality), Model Rule 5.5 (unauthorised practice) — we will not ship consumer-facing legal-advice products without a licensed attorney in the loop
  • ▸Sports / iGaming-adjacent: state-by-state wagering regulators (NJDGE, PA Gaming Control Board, MGM, UKGC, MGA) — we decline engagements in jurisdictions without a clear licensing path

Platform inventory & data-source catalogue

Every integration point named, every data licence read.

  • ▸Source-of-record systems: Salesforce GovCloud, MRI, Yardi, AppFolio, iManage, NetDocuments, Blackbaud RE NXT, Salesforce NPSP, John Deere Operations Center, Climate FieldView, Sportradar, Stats Perform
  • ▸Per-source metadata: refresh cadence, schema stability score, PII classification, data-licensing constraint (e.g. MLS feeds: display-only vs derived-products-allowed), authentication mechanism, rate limit
  • ▸Vendor-API health audit during discovery — we have walked away from engagements where the upstream API was too unstable to build production on (and we say so in the discovery report)
  • ▸Build-vs-buy call on each component: where an off-the-shelf product (e.g. CoreLogic data, Lexis-Westlaw citators, Bates-numbering tooling) is cheaper than rebuilding, we recommend it and integrate

Per-vertical compliance posture

FedRAMP, MLS data-licensing, ABA confidentiality, state gaming — read in full, encoded as code.

  • ▸SSP modules generated from infrastructure-as-code so the FedRAMP package matches the actual deployment — not a maintained-separately Word document
  • ▸MLS rule engine that scores every record on display-eligibility, attribution requirements, and derived-product allowance — enforcement happens at query time, not at ingest
  • ▸LegalTech ethical-wall enforcement at the search-index level — a model retrieving from a matter the user cannot access gets an empty result set, not a redacted one
  • ▸Sports / gaming-adjacent: per-jurisdiction policy module with age-gate, geo-fence, and licence-status checks before any user-facing prediction is rendered

Milestone-gated delivery

Four explicit stop-or-continue gates.

  • ▸Gate 1 (after discovery): do we agree on scope, regulatory boundary, target metric? If no, the engagement ends — and the discovery artefacts are yours
  • ▸Gate 2 (first integration live in staging): is the upstream data quality sufficient? If the real-world source is worse than the sample, we re-plan before committing further spend
  • ▸Gate 3 (pre-production promotion): validation metrics, fairness or compliance checks, regulator-facing artefact pack — all signed off
  • ▸Gate 4 (30 days live): actual user behaviour vs scoped use case; rollback or scale-up decision with a written memo

Risk register & engagements we decline

Where we will not take the work — and we say so up front.

  • ▸Regulatory clarity absent: unregulated crypto-adjacent products, unlicensed consumer legal-advice apps, public-safety decisioning where a human cannot remain in the loop
  • ▸Data-source instability: ingest-platform schemas that change quarterly without notice — we tell clients to wait two release cycles for the upstream to stabilise
  • ▸Mismatched regulatory and engineering timelines: when the regulator has not yet finalised guidance, building on the draft creates rework cost that exceeds the value of being early
  • ▸Co-development conflicts: we will not quietly recycle a competitor's bespoke work — reusable components are licensed transparently, exclusive components stay exclusive

What clients in novel verticals actually ask us

Honest answers on scoping, compliance, minimums, and when to wait.

Your vertical isn't on our industries list? Start with discovery.

A 2-3 week sprint that produces a regulatory map, platform inventory, data-source catalogue, sized engagement plan, and risk register. Fixed fee. Stop-or-continue gate at the end.

50+
Projects delivered
96%
Client satisfaction
2019
Founded — US, UAE & Pakistan