Building reliability layers for AI systems - constraints, gates, and deterministic pipelines.

18 · Bengaluru, India · LinkedIn · pranavharikumar2008@gmail.com

AI systems break in deployment for a specific reason: generation is treated as the end point rather than a component inside a larger architecture that should constrain and verify it. I'm focused on the layer between the model and the real world — deterministic pipelines, explicit validation gates, and systems that fail predictably rather than hallucinate silently. Both of my projects are different implementations of the same idea applied in different domains.

A session-based orchestration layer for LLM-driven customer service. Instead of letting the model answer freely, SVMP routes every query through a typed pipeline.

• Soft Debounce - 2.5s sliding window aggregates fragmented messages into Complete Thought Units before any AI logic runs. Reduces LLM API overhead by 40–60%
• Multi-Tenant Identity Frame - 3-dimensional coordinate system (tenantId · clientId · userId) hard-silos every interaction at the database level. No prompt engineering — enforced at the DB layer
• MUTEX Locking - Atomic state lock prevents race conditions and guarantees exactly-once processing
• Similarity Gate (≥0.75) - If confidence is below threshold, the system freezes and escalates to a human agent rather than hallucinating
• Intent Logic Fork - Transactional queries (order tracking, cancellations) bypass the LLM entirely and hit verified APIs directly. Eliminates transactional hallucination
• P90 internal latency < 5s post-debounce · 500+ adversarial red-team tests completed

A CLI tool that converts natural language tasks into deterministic, runnable Python programs. The LLM never generates executable logic — it only selects and wires components that already exist.

• Typed Node Registry - 23 nodes across data ingestion, transformation, storage, querying, and serving. Every edge is only legal if output type of node A matches input type of node B
• Structured Planner - LLM returns a JSON graph spec (nodes, edges, params). Plan normalizer handles format inconsistencies before validation
• Deterministic Validator - 5-check pipeline: node existence · type compatibility · DAG integrity via Kahn's algorithm · orphan detection · required parameter enforcement. Aborts with explicit errors before a single line emits

Student Marketing Venture - Scaled from $0 to $4,000 revenue in 3 months. Managed execution team, delivered 20+ client projects.

Labor Market Research (Independent) - 150+ structured interviews across Kerala. Identified wage asymmetry patterns between migrant and local workers. Drafting research paper.

Python JavaScript SQL MongoDB Pydantic Multi-tenant architectures Deterministic validation Constrained LLM planning

Graduating high school April 2026. Feel free to reach out for any enquiries.