Fine-tuned models, private deployments, and governed AI workflows for regulated institutions — built so your data stays in your environment and every decision keeps a human accountable.
Built for regulated industries
Banks, insurers, SACCOs, and pension funds cannot paste customer data into a public chatbot. We build the version that passes review.
Most enterprise AI projects stall in the same place: between a promising demo and a production system that compliance, security, and the board will approve. The model was never the hard part. Data residency, audit trails, evaluation evidence, and human sign-off are.
We are a senior engineering team that works on that gap. We customize models to your documents and vocabulary, deploy them inside your own cloud tenancy or data centre, and wrap every workflow in the controls a regulator expects to see.
We are honest about the tooling. Retrieval is the right answer for facts that change. Fine-tuning is the right answer for format, tone, classification, latency, cost, and for workloads where no data can leave your premises. We will tell you which one you need before you pay for either.
What we deliver
Each one is delivered as a scoped engagement with written acceptance criteria.
LoRA and QLoRA fine-tuning of open-weight models (Llama, Qwen, Mistral, Gemma) on your documents, plus distillation of frontier-model quality into small, cheap models. Includes dataset curation, PII scrubbing, and a before/after evaluation scorecard.
Production model serving inside your own AWS, Azure, or GCP tenancy, or on hardware in your data centre, including air-gapped setups. Sizing, cost-of-ownership model, access control, and audit logging included.
Permission-aware search and Q&A across policies, product manuals, circulars, and procedures. Every answer cites its source, and staff only retrieve what they are cleared to see.
Golden test sets, hallucination and bias testing, prompt-injection attacks, and a regression suite that runs on every model or prompt change. You get validation evidence, not a vendor promise.
AI inventory, use-case risk tiering, model cards, validation packs, and a draft AI policy mapped to NIST AI RMF, ISO/IEC 42001, and the NAIC model bulletin. We are your technical documentation partner; your compliance and legal teams keep sign-off.
Structured extraction from claims files, KYC packs, loan files, and underwriting submissions, with confidence thresholds, a human review queue, and integration into your core systems.
AI agents that work multi-step back-office processes end to end, with approval gates, full audit trails, and runbooks. Humans stay on every adverse decision.
Voice agents for inbound service and first notice of loss, live agent-assist, and automated QA scoring across 100% of calls instead of a sample.
Insurance
Claims and underwriting operations, where cycle time and leakage are measurable.
Capture first notice of loss by voice, chat, or email, structure it, and route it to the right handler with severity and complexity flags.
Turn a 200-page claims file into a sourced summary of liability, coverage, reserves, and open items that an adjuster can verify in minutes.
Surface inconsistencies, repeat patterns, and recovery opportunities for your investigators to review. Flags, not verdicts.
Extract and normalize broker submissions, check them against appetite, and queue the best risks first.
Give service staff and brokers instant, cited answers from policy wordings, endorsements, and product guides.
Draft replies, renewals, and endorsements for human approval, and handle routine status queries around the clock.
Banking
Document-heavy operations where analysts spend their day reading, checking, and drafting.
Check onboarding packs for completeness and consistency, extract the data, and hand the analyst a pre-reviewed file with exceptions highlighted.
Gather the context behind an alert, summarize it, and draft the case narrative for your compliance officer to review and file.
Spread financials, pull covenants and history, and produce a first-draft credit memo in your house format. The credit decision stays with your officers.
Classify, prioritize, and draft responses to complaints, with root-cause reporting for your conduct and risk teams.
Respectful, compliant reminders and payment-plan conversations over voice, SMS, and WhatsApp, escalating to staff when needed.
Cited answers from prudential guidelines, circulars, and your internal policies, so staff stop guessing and stop emailing compliance.
Governance
The questions your CISO, CRO, and regulator will ask, answered in the design.
AI drafts, recommends, and flags. Credit denials, claim rejections, and filings always carry a named human approver.
Deployed in your tenancy or data centre. Your data is never used to train shared models, and access follows your existing permissions.
Every prompt, source document, output, and approval is logged and exportable for internal audit and supervisory review.
Designed around the EU AI Act high-risk categories for credit scoring and life and health insurance pricing, the NAIC model bulletin on insurer AI use, GDPR, and the Kenya Data Protection Act 2019.
Straight talk
We fine-tune open-weight models that you can own and run yourself. Frontier models such as Claude are not generally available for customer fine-tuning, so where Claude is the right engine we customize it through retrieval, skills, and tool integration, accessed through your own cloud account. We do not train foundation models from scratch, we do not claim certifications we do not hold, and we do not let AI make adverse customer decisions without a human in the loop. Frameworks are referenced as design guidance; compliance sign-off remains with your own legal and risk functions.
How we engage
Each stage ends with a decision point and a deliverable you keep.
01
Two to four weeks. We review your data, systems, and risk posture, rank use cases by value and feasibility, and deliver a roadmap with a fine-tune vs. retrieval vs. prompt decision for each. Fixed fee.
02
Four to eight weeks. One use case, real data, inside your environment, measured against success criteria agreed up front, with an evaluation report your model-risk team can read.
03
Eight to sixteen weeks. Integration with core systems, security review, access control, monitoring, runbooks, and staff training. Phased rollout with a human-review period before any automation is widened.
04
Ongoing retainer. Drift and quality monitoring, regression testing on every model update, new use cases, and governance documentation kept current.
FAQ
What enterprise buyers ask us first
Usually RAG first. Retrieval is better when facts change and you need citations, which covers most policy, product, and regulatory Q&A. Fine-tuning earns its cost when you need a consistent output format or house style, high-volume classification, lower latency and cost from a small model, or a model that runs fully on-premises. Many production systems use both. Our assessment gives you a written recommendation per use case.
Open-weight models such as Llama, Qwen, Mistral, and Gemma, using LoRA and QLoRA, and we distil larger models into small ones you can run cheaply. You own the resulting weights. We do not offer fine-tuning of closed frontier models such as Claude; those are customized through retrieval, skills, and tools instead.
It does not have to. We deploy inside your own cloud tenancy or on your own hardware, including air-gapped environments. When a hosted frontier model is the right choice, it is accessed through your own enterprise cloud account under your data-processing terms, not ours.
Every system ships with an evaluation suite, documented test results, known limitations, and monitoring. We produce model cards and validation packs in the structure model-risk teams expect, and we work alongside your second line rather than around it.
We do not build it that way. AI prepares, summarizes, and recommends; a named person makes and owns every adverse decision. That is both the safer design and the direction regulators are taking.
Yes. We are based in Nairobi and work with clients worldwide. For Kenyan institutions we design around the Data Protection Act 2019, including its cross-border transfer rules, and can deploy on local infrastructure where residency is required.
In stages: a fixed-fee assessment, a fixed-scope pilot, then a production build and an optional monthly retainer. You can stop after any stage and keep the deliverables. We scope pricing after a briefing call, and size it to the institution, from SACCOs to tier-one banks.
A 45-minute briefing with a senior engineer. You leave with a view on feasibility, the right architecture, and what your risk team will need to see.