ArjunVerma
the moat is that I got the dawg in me
01About
- I've been writing code since I was twelve and still can't leave a system alone once I've seen inside it.
- These days it's mostly AI infrastructure, retrieval that actually retrieves and agents that finish the job.
- I once trained a 100M-parameter model on four billion tokens of Reddit Hinglish to see if it could text like my friends. It could, mostly.
- I build AI systems for a living, and take a few apart to see how they work.
I use Arch btw.
02Selected work, 14 items
Aug 2025 – present
- LLM agents write the Manim code, run it in a locked-down sandbox and fix their own bugs. RAG over the docs cut hallucinated code ~90%.
- A real-time collaborative video timeline on Yjs CRDTs. Basically Google Docs, but for editing video.
Attachments
- TypeScript
- Python
- Manim
- LangGraph
- LLaMa 3.1 8B
- RAG
- BullMQ
- Redis
- MongoDB
- FFmpeg
- Razorpay
Oct 2025 – present
- My own tokenizer, SFT + DPO, and it runs entirely in your browser at 100+ tokens/sec.
- Built the 4B-token corpus myself, and taught it tool use: it emits calc tags mid-generation and the runtime does the math.
Attachments
- PyTorch
- TPU
- BPE
- RoPE
- GQA
- SwiGLU
- SFT
- DPO
- ONNX
- transformers.js
- WASM
Jan 2026 – present
- 14 services on one machine, with an OpenClaw-style gateway I rebuilt myself, backed by 381 tests on real infra.
- Loads skills on demand, remembers past context with vector memory, and falls back to another LLM provider when one fails.
- TypeScript
- Node.js
- BullMQ
- Redis
- Postgres
- Next.js
- PI Agent
- mem0
- Skyvern
- Tavily
- PM2
Oct 2025 – Jan 2026
- The least clever thing I've built and the most used.
- 6 LLM calls run in parallel per resume (~6x faster), and recruiter mode screens up to 1,000 resumes per role, streamed live.
Attachments
- FastAPI
- Python
- MongoDB
- Next.js
- Vercel
- AWS EC2
- Docker
- SSE
- FingerprintJS
Aug 2026
- The LLM only flags evidence and code computes the odds, so nothing appears without a permalink.
- Pulls the full public history from Reddit archives through a Redis-backed queue, then maps communities, contacts and links.
Attachments
- Next.js
- TypeScript
- Redis
- Postgres
- LLM
Feb 2025 – Aug 2025
- Custom screen capture at 40-50 FPS (libraries manage 5-6) feeds a detector I trained on my own labels.
- The LLM reads the board as plain text and picks the next move, with multithreaded troop deployment in real time.
Attachments
- Python
- LLM
- Object Detection
- Windows API
- Multithreading
Feb 2026 – present
- It stays smooth with WebGL pieces on the page by keeping only 5-10 live previews at once.
- An importer turns any library into installable packages (2,057 so far), and I ported 128 Vue components to React.
Attachments
- Next.js 15
- React
- TypeScript
- Tailwind v4
- shadcn/ui
- Radix UI
- Shiki
- Playwright
Mar 2026
- I hand-wrote the crypto for each OS, and it works while Chrome is still open.
- pip-installable, with a CLI, a Python API and 7 output formats that plug straight into curl and wget.
Attachments
- Python
- Click
- PyCA cryptography
- DPAPI
- Keychain
- D-Bus Secret Service
- SQLite
Mar 2024 – Mar 2025
- Under 5MB, disguised as a system app, and it phones home through Google Sheets so IP changes don't lose it.
- The desktop build reaches files and the camera; the Android build reads contacts, location and notifications and hides its icon.
- Python
- React Native
- Node.js
- Google Sheets API
Feb 2026 – present
- It scrolls like a person, overshoots included, and doesn't break when LinkedIn redesigns its UI.
- 22 commands that answer in JSON with exit codes, so the agent calls it like any other function.
- Python
- Patchright
- Playwright
- Typer
- Rich
Apr 2026
- I reverse-engineered Codex's undocumented protocol from its CLI source, and it streams token by token like the real API.
- Handles the OAuth login and collapses a burst of expired requests into one token refresh. On npm as xypro.
Attachments
- Node.js
- Express
- OAuth PKCE
- SSE
- React
- Vite
- Tailwind
Feb 2026 – present
- X deletes tweets from the page as you scroll, so it catches them live, up to 5,000 per pull.
- 6 commands, no private APIs, and the same human-like layer as the LinkedIn CLI, verified against 5 bot-detection sites.
- Python
- Patchright
- Playwright
- Typer
- Rich
Jan 2024
- It cracked my own account on attempt 78.
- Built to spot sign-in flows with no rate limiting or lockout, the ones that are actually crackable.
- Python
Nov 2025 – Dec 2025
- Stateless Docker services and Redis queues hold p95 under 200ms, with a live map fed by 20+ government sources.
- OTP auth, MFA, CAPTCHA and role-based access for four user types, plus RAG over the training data on local quantized LLMs.
- React
- Node.js
- Docker
- Redis
- Nginx
- Leaflet
- RAG
- Quantized LLMs
03Writing
04Experience, 7 items
Sep 2026 – presentSan Francisco, California (Remote)
- Own Plutus, Atlan's AI for the finance team's month-end close. 1,000+ commits in 6 weeks, about 96% of the repo since I joined.
- Gave the agent hands: it edits bank lines, schedules and close tasks as you, but only behind a card you approve, with undo.
- Put a gate in front of the sandbox (concurrency caps, a waiting room, memory fences) so one heavy upload can't take prod down.
Jul 2026 – Sep 2026Remote
- Built the software factory: a Linear ticket goes in, one tested, human-approved pull request comes out. 710 tests, ~47K lines of TypeScript.
- The agent plans, a separate verifier re-runs the tests every round, and nothing can merge or touch main without a human.
- Built the context system that hands developers the requirements and strips who they're for. A salted-hash leak gate catches what the model misses.
May 2026 – Jul 2026Bengaluru, India (Onsite)
- Built Encube, which turns pharma SOPs into Duolingo-style lessons in 4 languages, on LangGraph, hybrid RAG and a Neo4j graph.
- Kept working memory under 8K tokens with 3-stage context compression. LLM spend down 54%.
- PDF parsing 150x faster with PyMuPDF, lesson generation 7x faster with async fan-out.
Feb 2026 – Apr 2026Bengaluru, India (Hybrid)
- Rebuilt production RAG into a 5-stage pipeline. Answer quality 4.12 → 8.93, latency 12s → under 5s.
- Picked the stack with data, not vibes: an LLM-as-judge harness, 33 experiments, 12 rerankers.
- Fine-tuned Qwen 3.5 (27B) into a tutor that stays in character. Rubric score 48% → 87%.
May 2025 – Jan 2026Chennai, India (Onsite, then Remote)
- Wired 50+ SliverC2 ops into an MCP layer to test whether an LLM can drive offensive-security tooling.
- Fine-tuned open-weight models with SFT + DPO until they stopped refusing legit security questions. Refusals 88% → 14%.
- Served them on one A100 at ~250 tokens/sec with vLLM and 4-bit AWQ. Only intern offered a remote extension.
Attachments
Feb 2025 – Mar 2025Remote (Paid)
- Made Express APIs serving 10K+ requests a day 40% faster, mostly by cutting query time 60% on the hot paths.
- Shipped JWT auth and role-based access across 5,000+ accounts.
Dec 2024 – Jan 2025New Delhi, India (Onsite, AICTE HQ)
- Built an offline-first message queue for Samvadini, the government's messenger. Failed deliveries down 85%.
- Hunted down N+1 queries on the busiest routes. P95 latency down 47%.
Links
05Extracurricular, 7 items
First place at Smart India Hackathon 2025, out of 72,000+ participants, the largest hackathon in the country. We built Prashikshak, a national-scale testing platform for the NDMA.
Links
Photos · 5
First place at Hacktivate 2025, against 1,200+ teams.
Links
Runner-up at the Level Super Minds Hackathon, out of 23,000+ participants.
Links
Runner-up at Brain Battles 2024, a national hackathon.
Links
Finalist at E-Summit 2025, a national hackathon.
Links
Finalist at Code of Phoenix 2025, a national hackathon.
Links
The highest grant the university awarded under its Capacity Building for Design and Entrepreneurship program, run by the E-Cell.
Links
06Contact
Building something cool? HMU! let’s talk. Email’s the fastest way to reach me.