Agentic AI
The pilot graveyard is an evaluation problem
Most stalled AI projects never defined what "correct" looks like. A 300-example golden set, built with the people who do the job, is the cheapest investment in your AI programme.
AI agents that run the work, not just the demo.
We design, build and operate agentic AI systems for banks, insurers, hospitals, universities and manufacturers — with the evaluation, observability and human approval gates that production demands.
40+production AI systems shipped
4regions served
12 wkstypical pilot-to-production
10.8 s
Classify incoming claim pack
14 pages · 3 document types detected · confidence 0.97
22.1 s
Extract policy, incident and invoice fields
31 fields · 2 flagged for low confidence
31.3 s
Retrieve policy terms and prior claims
Grounded on 4 internal sources · citations attached
43.4 s
Draft assessment and recommended decision
Eval score 0.92 · policy-rule checks passed
5—
Hand to human assessor
Decision above auto-approve threshold
Awaiting approval · every step logged and reproducibleApproval gate
Model-agnostic · traced · versioned promptsIllustrative run
What we build
Six capabilities, delivered as working systems rather than slide decks. Each one ships with evaluation sets, monitoring and a clear operating model.
Agentic workflows
Multi-step agents that plan, call tools, read your systems and complete tasks end-to-end — with approval gates where the risk warrants it.
Document intelligence
Turn forms, contracts, claims packs, lab reports and invoices into reviewed, structured data with field-level confidence and audit trails.
Enterprise knowledge & RAG
Grounded answers over policies, manuals and case history — hybrid retrieval, re-ranking and citations so every answer can be traced to a source.
LLMOps & evaluation
Offline eval sets, online monitoring, prompt and model versioning, drift and cost dashboards — the plumbing that keeps AI reliable after launch.
Small & private models
Fine-tuned small language models and on-premise deployments for teams that cannot send data out, or need predictable unit economics at scale.
AI governance & readiness
Model inventories, risk tiering, decision logs and board-level reporting aligned to MAS AI risk guidelines, PDPA, the India DPDP Act and the EU AI Act.
Recent case studies
Two of five anonymised engagements from the last 18 months.
A 19-day claims backlog cut to under 48 hours
The pressure
Motor and property claims arrived as mixed PDFs, photos and emails. Assessors spent most of their day re-keying data, and complaint volumes were rising with the backlog.
What we built
A claims intake agent that classifies the pack, extracts fields with confidence scores, retrieves the policy terms and drafts an assessment. Anything below threshold, or above a payout limit, routes to a human with the full trace attached.
87%of claims first-pass extracted
<48 hmedian cycle time
100%decisions logged and reproducible
Onboarding reviews that stopped missing the regulator's deadline
The pressure
Corporate onboarding needed documents from six jurisdictions reviewed against internal policy. Reviewers were inconsistent, and turnaround regularly breached the committed SLA.
What we built
A document-review copilot grounded on the bank's own policy manuals, with a checklist agent that drafts the review memo, cites every clause it relied on, and flags gaps for the analyst to confirm.
3.4×faster review turnaround
0SLA breaches in the first quarter
Every memocited to source policy
How we work
A fixed-scope path from a real business problem to a system your operations team runs. No open-ended discovery phases.
1
Assess
We map the workflow, the data, the risk tier and the economics. You get a build/no-build decision with a business case.
2 weeks
2
Governed pilot
A working agent on your data with an evaluation set, approval gates and a measured baseline — not a chatbot demo.
4–6 weeks
3
Deploy
Integration with your systems, hardening, observability, runbooks and training for the team that will own it.
4–6 weeks
4
Operate & improve
Managed service or handover. Weekly eval review, drift and cost monitoring, model upgrades without regressions.
ongoing
What we're telling boards this year
Short positions from our current engagements.
Cost & sovereignty
Small models are winning the second wave
For narrow, high-volume tasks, a fine-tuned 7–14B model on your own infrastructure now beats frontier APIs on cost, latency and data-residency — if you have the evaluation discipline to prove it.
Governance
Regulators are asking for traces, not policies
MAS guidance, the EU AI Act and India's DPDP rules converge on one practical demand: show the decision, the data it used and who approved it. Build the log before you build the agent.
Start with a working session
Bring one workflow that costs you time or money. In 45 minutes we'll tell you whether an agent can take it on, what it would need, and what it would cost — honestly.