HMC · 2025 · 8 months
Radiology AI, Hamad Medical
AI diagnostic assistant embedded in a live hospital radiology lab, in the clinical path rather than beside it.
- Clinics
- 1,000+
- Throughput
- 500+ scans/wk
- Sensitivity
- 94%
- Cost cut
- 78%
Available for engagements, Q3 2026
I build AI that knows when to refuse.
Hospitals, EU fintech, AML. Seven years shipping systems into places where a confident wrong answer costs more than no answer at all. Every claim cites its source or gets struck.
- Shape
- Scoped outcomes, 2 to 16 weeks. Never open-ended.
- Based
- New Delhi, IST. Works European hours.
- Start
- One 30 minute call. No deck, no discovery invoice.
- Leaves with
- Runbooks, tests, and a team that can hold it.
The gate
Pick your situation. It resolves or it declines.Most consultants say yes to everything and sort it out later. I would rather tell you up front when the answer is no, so pick whichever line sounds most like your week.
awaiting input
nothing is asserted until a situation is selected
nothing is asserted until a situation is selected
Evidence
Every engagement above cites one of theseFintech · 2025 · 6 months
Multi-agent credit rating
Five cooperating agents drafting rating reports behind a citation gate. The direct ancestor of Sentinel.
- Analyst time
- 70% cut
- Volume
- 250 to 1,250/yr
- Correctness
- 89%
- Value
- ~$900K/yr
Synaps · 2023 to 2025
Fraud and KYC ML
Detection across 50+ countries under EU AML5, FATF and OFAC. Fifteen production models on one MLflow pipeline.
- Precision
- 96%
- False pos.
- 23% lower
- Deploy
- 2wk to 3d
- Audits
- No findings
Nikxius / Sentinel · current
Eval-gated agent platform
Fourteen agents, five paying customers, six evaluators per run. Sentinel refuses to draft an AML narrative it cannot cite.
- Hallucination
- 0.5 to 0.95
- PII detect
- 0.7 to 1.0
- Tests
- 1,200+
- Run cost
- $3.25/cust/day
Capability index
49 of 49 shown| # | Service |
|---|
What I will not sell you
Cheaper for both of us if you know nowGeneric web and mobile builds
I can do it. So can a hundred cheaper people who will do it better because it is all they do. Hire them.
Foundation model training
Fine-tuning on your data, yes. Training a base model from scratch is a different discipline and a different budget.
24/7 managed operations
I build it, instrument it, and hand it over with runbooks. I am not a support desk and one person should never be your on-call.
Staff augmentation by the month
I take scoped outcomes with an end date. Open-ended seat-filling turns into the thing nobody owns.