┌────────────────────────────────────────────────────────────────┐
│ BSc Computer Science (AI) · University of Leeds · 2026 │
│ Systems Programming · Machine Learning · Quant Eng │
└────────────────────────────────────────────────────────────────┘
◈ About Me
name : Arjun Varma
interests:
- Low-latency systems & high-performance C++
- Machine learning engineering & LLM fine-tuning
- Quantitative / algorithmic trading infrastructure
- Distributed systems & cloud-native architecture
currently:
- Finishing Final Year Dissertation: LLM distillation for lecture understanding
- Building production-grade systems projects in C++ and Python
- Actively seeking: SWE, ML Engineering, or Quant Dev roles (2026)| Domain | What I'm Building |
|---|---|
| 🧠 AI / ML | LLM distillation, PEFT/LoRA fine-tuning, serverless ML pipelines |
| ⚙️ Systems | Lock-free data structures, cache-aware containers, smart pointer internals |
| 📈 Quant | Low-latency order matching, market data feed handlers, backtesting engines |
| ☁️ Cloud | Azure-native deployment, GitHub Actions CI/CD, containerised inference |
| Project | Stack | Headline Metric | |
|---|---|---|---|
| ⚡ | High-Performance UDP Multicast | C++ POSIX Lock-Free |
1.2M+ msg/sec · sub-100µs p99 latency |
| 📊 | Quant Trading Engine | C++ Multithreading |
1M+ order insertions/sec · <10µs latency |
| 🧠 | LLM Distillation Pipeline | PyTorch LoRA HuggingFace |
+59.5% BLEU · 7B→2B model compression |
| ☁️ | Serverless ML Pipeline | GitHub Actions Azure Python |
Event-driven · zero persistent servers |
| 🍔 | FoodVision Full-Stack ML App | TensorFlow Flask React Docker |
90% accuracy · EfficientNet on 10K+ images |
| 📡 | Real-Time Event Streaming | WebSockets Redis Protobuf |
Sub-ms delivery · SSR dashboard |
| 🔒 | Smart Pointer System | C++ Atomics ASan |
Thread-safe · lock-free reference counting |
| 📦 | Cache-Aware Containers | C++ SoA/AoS |
4-5× perf delta · bandwidth-bound analysis |
| 🖧 | Multithreaded File Server | C++ POSIX Thread Pool |
Bounded concurrency · backpressure under load |
◈ Industry Experience
Software Engineer Intern — Probe42 · Bangalore, India · Jul 2024 – Jun 2025
Financial data intelligence platform serving enterprise clients across India.
- Owned backend design for Project Mercury, a financial chatbot handling 500+ concurrent req/sec at 200ms median latency
- Built automated LLM evaluation pipelines, cutting manual validation effort by 75%
- Fine-tuned LayoutLMv3 on custom PDF datasets → 97% accuracy on financial field extraction
- Engineered load-testing infrastructure with Locust, measuring p50/p99 across distributed services
┌──────────────────────────────────────────────────────────────────────────┐
│ │
│ LLM Efficiency → distillation, quantisation, PEFT/LoRA │
│ ML Systems → inference optimisation, serving pipelines │
│ High-Frequency Trading → order book dynamics, ultra-low latency arch │
│ Distributed Systems → consensus, fault tolerance, event streaming │
│ Computer Architecture → cache hierarchy, memory models, atomics │
│ │
└──────────────────────────────────────────────────────────────────────────┘
◈ Engineering Notes
// Things I think about at 2am
// 1. A mutex acquisition can cost ~100ns. A cache miss costs ~100ns.
// They're the same problem wearing different hats.
// 2. The difference between p50 and p99 latency tells you everything
// about whether a system will hold under real load.
// 3. std::shared_ptr's control block is a heap allocation.
// make_shared merges it with the object — one alloc, better locality.
// 4. A 7-billion parameter model's "understanding" can be compressed
// into a 2-billion parameter model with the right distillation signal.
// 5. Lock-free doesn't mean wait-free.
// Progress guarantees matter more than the word "lock-free" on a slide.
