Built since Deloitte

Neydran

NSE stock screener and trading journal

Solo: architecture, backend, frontend, infrastructure

A live screener for Indian equities that shows its working. Instead of a score, every result carries the rules, numbers, and failed breakouts behind it.

• Ingests 5 years of NSE market data (3.3M+ price rows, 140 indices) into PostgreSQL and screens the Nifty 500 against an 8-rule trend template, recording every rule per stock.
• Terraform-managed Cloud Run stack with keyless GitHub WIF deploys. CI gates on Trivy, pip-audit, and gitleaks before Alembic migrations and a health check.
• Mandatory TOTP 2FA enforced at three layers (Next.js routes, FastAPI aal2 checks, Postgres row-level security). Chart queries cut from 1.28s to 0.09s with a composite index verified via EXPLAIN ANALYZE.
• NSE's bot filter returns 403 to cloud IPs, so ingestion runs as a scheduled job on a home machine, watched by a Cloud Run watchdog that alerts on missed trading days.

INKY Workspace Helper

Chrome extension for email-security rollouts

Solo, built for IT4Causes deployers

Deploying INKY into a nonprofit's Google Workspace means getting 200+ settings right, and one wrong checkbox can stop a client's mail. This extension turns that into one guided panel.

• Covers the full deployment: hosts, inbound gateway, and SMTP relay, with every IP range, header rule, and checkbox laid out in console order.
• Separate US and EU data-center profiles, each value sourced from INKY's documentation and dated so stale entries are easy to spot.
• Built while leading INKY rollouts across IT4Causes' nonprofit client tenants, from pilot testing through portfolio-wide rollout.

YOM

Bill-splitting app for iOS, Android, and web

Backend, AI receipt pipeline, and all infrastructure, with one co-developer

“You Owe Me”: split a bill and share it in under a minute, with no account needed for the first split or for the friend receiving it.

• Benchmarked several OpenAI vision models on ambiguous receipts and fed their output into a pure-integer ledger engine, so money math never depends on the model.
• Hardened the FastAPI backend for public launch: rate limiting, request-size limits, security headers, hashed one-time guest tokens, and app-store-compliant account deletion, backed by 160+ tests.
• Production on Cloud Run with Terraform-managed WIF, Trivy-gated CI/CD, and Cloudflare DNS. Resolved 5 launch-day incidents, including a blank web app caused by an untransformed import.meta in the static bundle.

AGENTS.md library

Coding-agent framework for research, build, and review

Solo

Engineering discipline for coding agents, in plain Markdown. One router file tells the agent exactly which few guides to read for the task in front of it.

• An ~80-line router classifies each task and hands the agent a 2–4 file reading list from 30+ guides covering core engineering rules and FastAPI, Terraform, React, database, ETL, and Ansible domains.
• Decision-council and adversarial-review protocols in which skeptic agents must document concrete attacks before work is marked done, runnable as parallel subagents or single-threaded.
• Neydran and YOM are standardized on its changelog, decision-record, and API conventions. It stays agent-agnostic: Claude Code, Cursor, or Codex.

bring

Workspace-first terminal UI for dev containers

Solo

Bring a project's dev container up, take it down, inspect it, and clean up after it, without remembering devcontainer or docker syntax.

• TypeScript CLI and full-screen TUI (Ink/React) published to npm across 6 releases through v0.3.0, managing any project's dev container by path, with ‑‑json output for scripting.
• Adds the stop and remove commands the upstream Dev Containers CLI lacks by locating containers through devcontainer labels and managing them via Docker, spawning every command without a shell.
• ~300 tests across a hermetic suite (fake binaries, temp state) and a real-lifecycle integration suite against Docker. Dotfiles profiles bring new containers up preconfigured.

Multi-tenant RAG

Retrieval research, Georgia Tech D2I lab

Retrieval pipeline and model backends, with a lab teammate

A reproducible multi-hop RAG baseline on the MuSiQue benchmark with per-user indexes, built to find out exactly where a naive pipeline loses answers.

• Ran a step-by-step ablation (indexing granularity, context format, system prompt, generator), measuring every change on the benchmark and reverting the ones that regressed.
• Rebuilt retrieval around a sentence-level embedding index (Qwen3-Embedding-8B) that maps each hit back to its parent chunk before generation.
• Instrumented per-stage latency and added pluggable vLLM and OpenAI generator backends behind one config-driven pipeline.

FitAI

Adaptive AI fitness coach and HCI study, Georgia Tech CS 6641

Lead engineer on a three-person team

An LLM coach that rewrites tomorrow's workout from how sore you are today, built to test whether adaptive plans actually feel more personal.

• Built the full stack: a Next.js app, a FastAPI coach service that checks the Supabase JWT on every route, and Postgres with row-level security across 19 migrations.
• Tap the muscles that are sore and GPT-4o mini proposes tomorrow's swaps through a forced tool call. A validator rejects any exercise outside the 1,273-entry catalog before the plan changes.
• Ran a within-subjects crossover study comparing adaptive AI plans against static ones, from protocol design through analysis.