ArjunVerma

I am a curious Engineer at heart. Been at this since I was twelve.

Find me elsewhere

GitHub↗

About

I build software I'd want to use.
Systems, interfaces, and the occasional weekend experiment.
I write to think clearly, and I ship to learn faster.

I use Arch btw.

02Selected work^[13]

You describe a concept in plain English and Stitcher writes the Manim code, runs it, and renders an actual animated video. Around 8K signups and 50+ people using it daily.
A crew of LLM agents handles codegen across 10+ providers and debugs its own mistakes in a retry loop. RAG over the Manim docs killed about 90% of the hallucinated code.
Every AI written script runs inside a locked down sandbox (bubblewrap + seccomp) so untrusted generated code can never touch the host machine.
Built a real time collaborative timeline editor with Yjs CRDTs over WebSockets. Basically Google Docs, but for editing video.
Payments are race safe: the Razorpay verify call and the webhook can both fire and you still get charged exactly once, thanks to an atomic Mongo update.

Attachments

Demo video↗

try it live!↗

TypeScriptPythonManimLangGraphLLaMa 3.1 8BRAGBullMQRedisMongoDBFFmpegRazorpay

Trained a 99M parameter transformer from scratch, same architecture family as Llama and Mistral (RoPE, grouped query attention, SwiGLU, RMSNorm).
Wrote my own byte level BPE tokenizer with a 32K vocab and built the full 4 billion token training corpus myself, deduping millions of near duplicate docs with a Bloom filter.
Full post training stack: supervised fine tuning, then DPO, with an LLM as a judge harness scoring relevance, persona and coherence on a held out set.
The whole thing runs client side in the browser (PyTorch to ONNX to WebAssembly, int8 quantized) at 100+ tokens/sec with nothing running on a server.
Taught it to use tools mid generation: it emits plain text calc tags the runtime catches and solves, so it does arithmetic cleanly and gets ~56% on harder word problems.

PyTorchTPUBPERoPEGQASwiGLUSFTDPOONNXtransformers.jsWASM

Built a 14 service agent system from scratch (gateway, main agent, 5 connectors, cron, a human in the loop dashboard) that runs 24/7 on a single machine.
Read the OpenClaw architecture and implemented its gateway myself: FIFO ticket dispatch over Redis, heartbeat based eviction, and recovery of in flight work if a worker dies mid task.
The core agent loads skills on demand, remembers past context with a vector memory (mem0), and can switch between LLM providers at runtime with automatic fallback when one fails.
Gave it real hands on the web with Skyvern for browser automation and Tavily for live research, so it can actually go do tasks instead of just talking about them.
Backed by a 381 test monorepo with real infrastructure integration tests spun up through Testcontainers, not mocks.

TypeScriptNode.jsBullMQRedisPostgresNext.jsPI Agentmem0SkyvernTavilyPM2

03Writing

23.06.2026·9 min read

Experience

• Built a unified multimodal ingestion layer across regulatory inputs (PDFs, DOCX, images, audio) using Docling and VLMs, hitting 94% structured field extraction accuracy. • Shipped a 3 stage context compression middleware that keeps working memory under 8K tokens, cutting per session LLM spend by 54%. • Owned the SAE completeness module for the CDSCO India AI Hackathon 2025, a 4 layer consistency model where every finding cites its source regulatory doc (NDCTR 2019, 2014 Pre screening Checklist, ICMR CECHR, Schedule Y). • Built the entire FastAPI backend (35+ endpoints) and the LangGraph pipeline for Encube Learning Solution, fanning out across 8 generation skills to turn pharma SOPs into Duolingo style lessons in 4 languages. • Got PDF parsing 150x faster swapping docling for PyMuPDF, and pushed lesson generation 7x faster going parallel with asyncio.gather.

Attachments

Offer Letter

• Redesigned production RAG into a 5 stage pipeline (HyDE, topic filter, wide retrieval, rerank, final pick), lifting eval quality from 4.12 to 8.93 and dropping latency from 12s to under 5s. • Built an LLM as judge eval harness and ran 33 experiments benchmarking 12 rerankers, 3 embedding models and 5 pipeline architectures. Big finding was that embedding choice moves the needle way more than the reranker. • Curated an 85K tutorial dataset, cleaned it with a recursive MathML to LaTeX converter and a 4 stage dedup pipeline with number signature and correct answer guards so near duplicate questions with flipped meanings don't get falsely merged. • Fine tuned Qwen 3.5 (27B) into a persona conditioned tutor on 85K SFT + 30K preference pairs, taking the rubric score from 48% to 87%.

Attachments

Offer Letter

• Built a full stack C2 controller for a Red Teaming framework using gRPC with mTLS and Protocol Buffers (cut data size by 40%), wired JWT + RBAC for 50+ operators. Sole intern to get a remote continuation offer with a 28% stipend hike. • Designed a centralized operator audit system capturing IP, device fingerprint, geolocation and session telemetry, processing 100K+ events and detecting unauthorized access with 99% accuracy. • Deployed a scalable C2 server with AWS EC2 + MongoDB cluster + Docker network isolation, secured via Nginx reverse proxy, Cloudflare Full Strict HTTPS, X.509 internal mTLS and DNS based redirectors that obfuscate the real C2 IP. • Shipped an ML based request filter that blocked 95%+ of malicious traffic and cut false positives by 20%, plus a RAG based CVE intelligence system mapping host fingerprints to known vulnerabilities in real time. • Built an agentic red teaming MCP compatible system powered by Qwen 2.5 (32B) fine tuned with LoRA + RLHF, with grounding and episodic memory for autonomous operations.

Attachments

Offer Letter Letter of Recommendation

Attachments

CyStar Completion Certificate.pdfAttachment

• Prototyped a service that runs daily unit tests on products in development, cutting bug identification and fix time for the team. • Built and debugged RESTful APIs in Express.js handling 10,000+ requests a day, with 40% better response times and tighter auth. • Implemented secure JWT auth and RBAC, managing 5,000+ user accounts safely. • Refactored database queries dropping execution time by 60% on high traffic endpoints, and built n8n automations + MCP servers with tool calling, state management and guardrails for hallucination control in agentic AI systems.

• Worked with the core team behind Samvadini, the Government of India's multilingual messaging platform. • Shipped an offline first message queue with IndexedDB sync, dropping failed message deliveries by 85%, plus a real time messaging pipeline on WebSockets + Redis pub/sub with under 100ms latency. • Built a contact syncing and fetching feature for Samvadini, cutting API calls for newly fetched contacts by 30% through an optimized algorithm. • Programmed and linked 20+ new screens in React Native to the backend, improving page load performance by 75% (3s to 0.75s). • Re architected backend hot paths with request tracing and N+1 elimination, bringing P95 API latency down by 47%, and executed MongoDB schema migrations on a live system with 100K+ documents without service interruption.

Attachments

Offer Letter Samvadini Messenger

Extracurricular

02Selected work^[13]

You describe a concept in plain English and Stitcher writes the Manim code, runs it, and renders an actual animated video. Around 8K signups and 50+ people using it daily.
A crew of LLM agents handles codegen across 10+ providers and debugs its own mistakes in a retry loop. RAG over the Manim docs killed about 90% of the hallucinated code.
Every AI written script runs inside a locked down sandbox (bubblewrap + seccomp) so untrusted generated code can never touch the host machine.
Built a real time collaborative timeline editor with Yjs CRDTs over WebSockets. Basically Google Docs, but for editing video.
Payments are race safe: the Razorpay verify call and the webhook can both fire and you still get charged exactly once, thanks to an atomic Mongo update.

Attachments

Demo video↗

try it live!↗

TypeScriptPythonManimLangGraphLLaMa 3.1 8BRAGBullMQRedisMongoDBFFmpegRazorpay

Trained a 99M parameter transformer from scratch, same architecture family as Llama and Mistral (RoPE, grouped query attention, SwiGLU, RMSNorm).
Wrote my own byte level BPE tokenizer with a 32K vocab and built the full 4 billion token training corpus myself, deduping millions of near duplicate docs with a Bloom filter.
Full post training stack: supervised fine tuning, then DPO, with an LLM as a judge harness scoring relevance, persona and coherence on a held out set.
The whole thing runs client side in the browser (PyTorch to ONNX to WebAssembly, int8 quantized) at 100+ tokens/sec with nothing running on a server.
Taught it to use tools mid generation: it emits plain text calc tags the runtime catches and solves, so it does arithmetic cleanly and gets ~56% on harder word problems.

PyTorchTPUBPERoPEGQASwiGLUSFTDPOONNXtransformers.jsWASM

Built a 14 service agent system from scratch (gateway, main agent, 5 connectors, cron, a human in the loop dashboard) that runs 24/7 on a single machine.
Read the OpenClaw architecture and implemented its gateway myself: FIFO ticket dispatch over Redis, heartbeat based eviction, and recovery of in flight work if a worker dies mid task.
The core agent loads skills on demand, remembers past context with a vector memory (mem0), and can switch between LLM providers at runtime with automatic fallback when one fails.
Gave it real hands on the web with Skyvern for browser automation and Tavily for live research, so it can actually go do tasks instead of just talking about them.
Backed by a 381 test monorepo with real infrastructure integration tests spun up through Testcontainers, not mocks.