Diagnostic Engine: Registry-Driven Knee Symptom Reasoning
A writeup on a bounded knee symptom engine built around registry-defined evidence, safety-first escalation, deterministic scoring, and an inspectable session ledger.
LLM Systems · Memory · Reliability
Governed systems for memory, execution, and reliability.
I build Governed AI infrastructure with a focus on reliability and deterministic execution. I work on research and systems projects around LLM workflow compilation, memory layers, and multi-tenant governance. I also write about system design, reliability, and AI architecture.
About me
About me.
I’m Pranav, an 18-year-old student from Bangalore interested in systems, startups, and how messy real-world work can be made more reliable.
Most of what I build sits around one idea: AI is useful, but it needs structure around it. I’ve worked on projects like SVMP Systems, a reliability-focused customer support system for LLM workflows, PlanCompiler, a research project on deterministic code compilation, and a governed memory layer for multi-agent systems. The common thread is trying to make software behave more predictably when things get complex.
More recently, that same thread has also shown up in Diagnostic Engine, a bounded knee symptom reasoning project built around registries, deterministic scoring, and safety-first escalation.
Currently working on SVMP Systems, which is my own small research/product lab (teamed with a few friends), focussing mainly on implementing governed architecture to real world use cases.
I’m still early, still figuring things out, and still learning in public. This site is where I keep the work that feels important enough to explain properly: projects, research notes, essays, and the occasional unfinished idea that might become something.
Latest Writing
Latest writing.
The Basilisk Is the Race, Not the Model.
A piece on why the danger in AI is not one future model or one person, but the race condition that makes governed architecture the realistic safety frontier.
Read blog →Deterministic Compilation for Structured LLM Workflows
A compiler-first approach to workflow generation where the model plans inside a typed node registry and execution only starts after static validation. 8x cheaper than gpt-4.1 and 25 point delta in first pass success over 300 task benchmark set.
Governed Memory Layer for Multi-Agent Shared State
A systems writeup on governed promotion, append-only events, pending ambiguity, and how canonical shared state should be earned rather than appended.
Read writeup →Selected Projects
Selected projects.
PlanCompiler
Typed, deterministic workflow compilation with validation ahead of execution. Beat gpt-4.1 and claude sonnet 4.6 with a 20 point delta in first pass success.
Governed Memory Layer
Ledger-backed shared memory for agents with pending-state deferral and deterministic projection.
Diagnostic Engine
Knee-only symptom reasoning engine with registry-driven evidence mapping, safety-first escalation, and an inspectable session ledger.
SVMP Governance Layer
Multi-tenant customer-service reliability layer for high-risk domains. contains inbuilt multi-tenant isolation, complete auditability, governance gates and a soft debounce architecture.
Contact
Open to research, systems work, and collaboration.
If you're building AI infrastructure or reliability tooling, feel free to reach out directly.
Writing updates
Follow the blog.
Get new architecture notes on governed AI systems, memory, reliability, and deterministic execution.