Enterprise AI delivery
Generative AI, governed data access, full-stack applications, deployment, and production operations.
Enterprise AI · Data platforms · Production systems
I lead enterprise AI from strategy to adoption and reliable operations.
I’m Tryambak Kaushik, an engineering leader with 15+ years across software, AI, cloud, and data platforms. I connect business needs to architecture, implementation, and reliable operation.
What I bring
From business need to reliable operation.
I work across requirements, architecture, implementation, and operations—and remain accountable for what happens after launch.
Cloud & data platforms
GCP, AWS, Azure, Kubernetes, Spark, Kafka, BigQuery, and enterprise data pipelines.
Multidisciplinary leadership
Aligning technical teams, business stakeholders, executives, external partners, and delivery constraints.
Selected work
Two practical tests of safer AI deployment.
Both prototypes ask a simple question: can AI do useful work while people can still review important decisions, understand the safeguards, and see clearly when something fails?
01
Robot Release Gate
A software update passes ordinary checks. The robot still misses the pickup.
A deterministic conveyor experiment exposes the failure conventional software metrics overlook: 180 ms of added decision latency changes the physical outcome, so the release is blocked.
Pass Model equivalencePass ThroughputBlock Task completion
Current boundary: browser-based timing experiment; MuJoCo physics validation is the next step.
02
SQL Execution Gate
An AI agent that must show its work before it can touch data.
The deployed assistant separates SQL planning from execution, exposes generated SQL for a human decision, blocks standard SQL writes mechanically, and fails closed on execution errors.
Question→Planner→Human approval→Read-only executor
Writing & technical studies
Ideas explained through mechanisms and evidence.
These studies were originally published through Aggregate Intellect, founded in Toronto in 2018, as part of its A.I. Socratic Circles (AISC). The original pages were later removed, so I now host edited versions here and link to the source reports.
Sequential recommendations
How BERT4Rec learns from both sides of a sequence
Architecture, masked-item training, evaluation, and what the reported gains do—and do not—prove.
Read case study → Reinforcement learningRNN versus Transformer policies for a trading experiment
An educational comparison with exact reported results and a frank methodology audit.
Read case study → Semi-supervised learningLearning with GANs when labels are scarce
A practical walkthrough of semi-supervised generative adversarial networks.
Read on KDnuggets ↗Let’s compare notes
Building consequential AI platforms or production systems?
I’m interested in the problems that appear after the model works: deployment safety, runtime reliability, evaluation, and accountable delivery.