18 · Bangalore · Builder & Researcher

Pranav Harikumar

LLM systems, deterministic compilers

I build things and write about them. Most of my recent work is about making AI systems more reliable; less prompt engineering, more structural guarantees. I also do field research on labour markets when I'm not at a keyboard.

LLM Reliability Deterministic Compilation Multi-Tenant Systems
See the Work Read the Paper

A bit about me

I like building systems that fail predictably instead of silently.

I'm 18, based in Bangalore, finishing school in April 2026. Most of my time right now goes into LLM systems work, specifically the question of how you build systems that fail in predictable, debuggable ways rather than silently misbehaving. That's what SVMP and the compiler paper came out of.

On the side, I've been doing independent field research on wage asymmetry between migrant and local workers in Kerala, ran survey with 150 respondents, working toward a paper.

Availale for any opportunities in LLM systems, deterministic compilers, or feedback.

Education

CBSE (PCMB), Sri Chaitanya Techno School — April 2026

Working on

LLM reliability

Location

Bangalore, India

Stack

Python · MongoDB · FastAPI

Open to

Fellowships · Research internships · Early-stage roles

Links

GitHub ↗  ·  LinkedIn ↗

Projects

Research Build · 2026

LLM Code Graph Compiler

Deterministic Workflow-to-Code

Separates LLM planning from execution entirely. The model selects nodes from a pre-verified typed registry and outputs a JSON plan. A seven-check static validator enforces correctness before any code runs; no repair loops, no runtime chaining, no silent failures. The evaluation surfaced a "constraint evasion" failure mode: leave one surface unconstrained and the planner always finds it. That turned out to be the most interesting result.

Python DAG Execution Static Validation CLI Tooling
278/300 Compiled runs
~202/300 GPT-4.1 baseline
183/300 Claude Sonnet 4.6 baseline
7 Static checks
02

Field Research · 2026

Wage Asymmetry Study - Kerala Labour Markets

150+ structured interviews across Kerala on the wage gap between inter state migrant workers and local labourers. Designed the interview protocol, ran the field data collection, writing it up as an economics paper. Short version: the gap is real, persistent, and institutional.

Field Research Labour Economics Structured Interviews
150+ interviews Kerala, India Paper in progress
03

Operations · 2025–Present

Niyomilan — Pre-Launch Labour Marketplace

Working on ops and growth for a marketplace built for daily wage workers. Field driven GTM work and early operational design. The Kerala research feeds directly into how I think about the problem here.

Operations GTM Labour Markets
Pre-launch
04

Venture · 2024-2025

Student Marketing Venture

Co-ran a content and marketing operation serving international clients out of school. Handled clients, revisions, payments, and delivery across 20+ projects. ~₹3.5L in three months with a small team. Learned a lot about what "ops" actually means.

Client Management Content Operations
₹3.5L · 3 months
05

Side Project · 2026

Matiks Breaker

Reverse-engineered the Matiks daily puzzle flow via its GraphQL endpoint, converted raw 7×7 state into structured constraints, built a permutation-based backtracking solver. The work later informed an independent public fan site for daily verified solutions.

Python GraphQL Backtracking

Get in Touch

Open to fellowships,
research & early-stage work

If you're working on something interesting, reach out.

Say hello