Zubin Deepak Rajasekar
Data & AI · London
Nine years across data, analytics and applied AI, including seven years at Deloitte and two years building AI products at NeuralHue. I've worked across public sector, healthcare, legal and professional services, designing systems, shaping requirements, evaluating what works, and taking ideas from problem definition through to delivery.
Profile.
AI delivery needs five things. I have done each one.
- 01
Find the problem worth solving.
Seven years translating US government and healthcare objectives into delivered systems.
- 02
Decide what to build.
Scoped AI systems across legal knowledge, talent intelligence and ESG assurance.
- 03
Know whether it works.
A 30 query adversarial kit, six scoring axes, noise measured at plus or minus one.
- 04
Know when to stop.
Killed a designed feature when 16 calibration tests produced no defensible threshold.
- 05
Land it.
Production platform delivered with source and docs. Client owned, extended beyond scope.
Work.
Live demo
FirmMemory
Governed retrieval over a law firm's matter archive. It refuses when the evidence is thin.
Delivered
Semantic talent intelligence
Search that matches people by what they have done, not what words they used.
Advisory
ESG governance assurance
A disclosure assurance engine where the legal boundary is structural, not a policy.
In progress
Precedent
Checking complaint decisions against how the Ombudsman has actually decided comparable cases.
Builds.
Writing.
22 August 2026Five questions between the business case and the build
Every AI programme starts with an objective and ends with a system. These are the five questions in between.
Read at neuralhue.com →13 August 2026Your archive knows things no single matter file knows
The answer to a panel application, a pitch, or a sector build-out may already be in the firm, spread across matters, years and people. No single document contains it. The archive does.
Read at neuralhue.com →25 June 2026The most useful answer a legal AI can give is "the firm hasn't done this."
Ask a legal AI a question it cannot answer, and most will answer anyway. The capability that protects a firm is the quieter one: knowing when it has nothing worth saying, and saying so.
Read at neuralhue.com →21 May 2026What we learned generating a 700-document synthetic legal corpus
The first thing a Knowledge Director asks when I show them FirmMemory is whether the demo is running on their data. The answer is nobody's. This note is about why a fake law firm was the right way to build a real product.
Read at neuralhue.com →Get in touch.
If you are building AI systems in a setting where the answers have to be defensible, email is the best way to reach me.
[email protected]