abhinav.phi

NSUT Delhi - B.Tech CSE '28 - New Delhi, India

I make software that holds up when things break.

Engineer by training, builder by habit. I like the parts of systems nobody photographs - the ledgers, the evals, the query plans. Currently deep in agentic AI and backend infrastructure.

NOW - building: reflex v2 · learning: Go · reading: Designing Data-Intensive Applications open to: internships & collaborations - onsite or remote · abhinav.phi15@gmail.com

01 - Selected Work

A few things I've built.

NIDSPython · XGBoost · WebSockets

A network intrusion detection system - 78 flow features per event, XGBoost separating 13 attack classes, live alerts streamed over WebSockets.

GitHub ↗
SutraFastAPI · Python · AWS EC2

A guardrailed message-composition service for a merchant AI assistant - validation gates, three-tier fallback chain, deployed on AWS.

GitHub ↗
optyraPython · Telegram API

GSoC issue monitor - tiered GitHub polling, rule-based scoring, AI two-line summaries, Telegram alerts.

GitHub ↗
ARGUS-LEOTypeScript

LEO satellite tracking and analysis tooling - my work on the data pipeline and visualization layer.

GitHub ↗
TensorGuard-LitePython · PyTorch

Sovereign AI auditor - weight fingerprints over 10+ GB models, PCA lineage clustering, GPU memory tuning.

GitHub ↗
NeuroAd EngineTypeScript · RL

A cognitive ad-testing environment - simulated audiences, RL-driven creative selection.

GitHub ↗

02 - Capabilities

What I do, in quiet detail.

Backend systems

REST APIs and event-driven services in Python - FastAPI, PostgreSQL (schema design, RPC, triggers, query optimization), Redis, WebSockets. The unglamorous layers where correctness lives.

Agentic AI

LLM systems with deterministic guardrails - rules-first diagnosis, expected-value decision ranking, evaluation design, and fallback chains. The model is the easy ten percent; the system around it is the work.

Optimization

Constrained scheduling and decision problems - OR-Tools CP-SAT under hard time budgets, deterministic fixed-seed simulation, reproducible evaluation pipelines.

Competitive Programming

Codeforces Expert (1600+), LeetCode 1650+, 500+ problems solved in C++ - daily problem-solving at contest-grade speed. The problem-solving core that everything above sits on.

Quant Research

WorldQuant BRAIN, Gold level - alpha signal ideas built and iterated through real backtesting discipline. Pattern-hunting in noisy data is the same muscle my detection and eval work trains.

Ship & operate

Docker, CI/CD with real test suites, AWS EC2 deploys, Linux as a daily driver. Code that passes review, and systems that survive contact with users.

03 - Credentials

Proof, not promises.

Open Source

50+ issues resolved across the open-source ecosystem

Issue triage, reproduction, fixes and test coverage contributed across Apache, Kubernetes, LLVM, ArduPilot, Django, Linux, Blender, Dart, Ceph and Checkstyle - reading large unfamiliar codebases is the core skill, and it runs deep.

Apache StormCrawler

Crawler performance + WARC integrity

Replaced a quadratic metadata-append path with a single bulk write on the hot crawling loop, and hardened WARC record serialization against CR/LF injection in metadata fields.

Kubernetes

Scheduler + admission docs and fixes

Worked through the scheduler and admission-control surface - documented MutatingAdmissionPolicy behavior, fixed stale anchors and API-reference regressions across the cluster-administration pages.

LLVM

CMake host-triple handling

Special-cased HarmonyOS hosts in get_host_triple() so cross-builds configure correctly instead of falling through to generic triples.

ArduPilot

AP_Notify LED driver extension

Added a StandardSwappedRedGreen LED override to the notification backend - boards with swapped polarity now get correct status colors without hardware changes.

Django · Linux · Blender · Dart · Ceph · Checkstyle

Triage across six more ecosystems

Issue reproduction and targeted fixes - from ORM edge cases and kernel-side behavior reports to Blender Python API quirks, Dart tooling, Ceph operator manifests and Checkstyle rule regressions.

04 - Terminal

$ whoami abhinav - b.tech cse, nsut delhi · class of 2028   $ ls ~/work reflex/ nids/ rail-bloc/ sutra/ apas/ +8 more   $ cat principles.txt guardrails before autonomy measure, then claim integrity at the data layer   $ uptime 3 Apache PRs merged · Codeforces 1600+ · WorldQuant BRAIN Gold $ exit session closed.