I'm Arin. I build systems that survive production.
Software Engineer at Saviynt Identity governance across 300+ enterprise tenants on AWS, Azure and GovCloud. Before that, four years of Ocean Engineering at IIT Madras (JEE top 0.2% of 1.1M). Heavy mathematics, computational modeling, physics. It trained me to reason about complex systems under extreme constraints. The systems just ended up being software rather than ships.
The 21-hour job that hooked me
Large companies must continuously prove that no employee holds conflicting permissions. Creating and approving the same payment, for example. Running this check is called segregation-of-duties evaluation, and one customer's job ran 21 hours on a 64GB machine. It re-examined the same organization hierarchy separately for every rule, thousands of times over. I replaced it with a single pass over the hierarchy plus bit-level permission comparison, where one processor instruction evaluates 64 people at once.
The remaining minutes are almost entirely database writes. Persisting the violation records is now the job, where evaluating them used to be. Making compute negligible didn't eliminate the bottleneck. It moved it somewhere cheaper. That migration is the shape of most performance work at scale.
That's the work I keep chasing: the right data structure collapsing a problem by two orders of magnitude.
What I work on
LICENSE INTELLIGENCE · LATEST
Companies pay SAP per user by access level, and overpay for dormant or over-privileged accounts. Most recently I built the inference engine that finds this waste over access data. Any number of accounts or identities, with dollar savings attached. It acts on what it finds, creating removal tasks with records that survive the nightly data refresh. Full story.
EXPORT PIPELINE
Streams reports of any size. 15M+ rows proven. Formatted Excel inside ZIPs directly to S3. Memory stays flat at 7MB. Paginated reads → 50-row sliding window → S3 multipart → streaming ZIP. Full breakdown in this post.
AGENT SAFETY
When AI assistants hold tools, data access, and credentials, and can call other assistants or share credentials with them, dangerous permission combinations reappear with no person in the loop. I looked for existing work on this problem, found none, and built the evaluation engine. Five risk patterns, deterministic checks, findings tracked to resolution. Full model · Technical report (PDF).
AUTOMATED FRAUD REVIEW
Software agents that investigate emergency-administrator logs on their own, within strict budgets on steps and time. Every log entry cited as evidence is re-checked against the source data; uncertain findings go to a human reviewer. New fraud patterns ship as configuration, not code changes.
SECURITY INITIATIVES
Twice asked to join temporary company-wide security teams: categories of injection and access-control flaws across services, encrypted internal communication, test coverage from 30% to 80%.
AFTER HOURS · AI SYSTEMS
Nostos, a KV-cache-aware LLM router (radix-tree prefix affinity, write-up); ArinLM, a GPT built by hand with no model shortcuts; the same transformer in dependency-free C++ (gpt2-cpp, tested); and again in raw CUDA (in progress). Why three times.
How I got here without a CS degree
My department's timetable left no room for Data Structures, OS, or Databases. So I trained in programming contests against students who had taken those courses, and stayed until my rankings matched theirs. Four years of evenings after naval architecture labs. Progress was slow and nonlinear.
- LeetCode Guardian 2100+ — ranks 248, 617, 630, 880 / ~45K
- Codeforces Expert 1602 — top 0.3% globally
- Meta Hacker Cup 2024 — Round 2 (National 424, Global 1903)
I learn the same way in ML: by rebuilding. A GPT by hand in PyTorch, then in dependency-free C++, then in raw GPU code; helmet detection with Faster R-CNN and YOLOv8; a Rubik's cube solver built on heuristic search with pattern databases.
Beyond engineering
Team Abhiyaan (IITM autonomous vehicles). I led external relations, pitched directly to India's Commerce Minister, grew reach 125%. Athletics: football city runner-up, Villarreal CF Academy, district champion; chess district runner-up; cricket district all-rounder.
What I'm working toward
I can build systems that work under extreme constraints. I've done it repeatedly at scale. But I've done it by inferring principles from operating production systems under pressure. I know that my exactly-once pipeline is correct. I want the theory that explains why in the general case.
The current list (consensus, storage engines, verification, and serving intelligence cheaply) lives on the homepage, next to the production experience behind each item.