machine learning · quantitative finance · low-level systems
classSrisailesh:
def__init__(self):
self.role="Engineer & researcher"self.works_on= ["machine learning", "quantitative finance", "systems"]
self.languages= ["Python", "C++", "JavaScript"]
self.principle="build it, then try to break it before reality does"defwhat_i_care_about(self):
return"code that's fast, results that are honest, and ideas that survive testing"- 🤖 Machine learning & AI — reinforcement learning, neuro-symbolic reasoning, applied ML on messy real-world data
- 🏗️ AI infrastructure — inference systems: paged KV-caches, continuous batching, and serving stacks built from scratch
- 📈 Quantitative finance — market microstructure, execution, and systematic strategies
- 🗳️ Distributed systems — Raft consensus from scratch, linearizability checking, and Jepsen-style fault injection
- ⚙️ Systems & software — low-latency C++, simulation engines, and full-stack apps
| Project | Domain | What it is |
|---|---|---|
| 🗳️ raftkv | Distributed systems / Go | A distributed, linearizable KV store on a from-scratch Raft (no consensus libraries): fsync WAL, lease reads, exactly-once client sessions, and dynamic sharding with live migration — verified by a from-scratch linearizability checker that CI runs against 1000 randomized fault schedules per push. Found-and-fixed bugs written up with root causes in BUGS.md. |
| 🤖 tinyserve | AI infra / Systems | A from-scratch LLM inference server (vLLM-lite): manual Qwen2 forward pass over a paged KV-cache, continuous batching with preemption, an OpenAI-compatible streaming API, and speculative decoding with exact verification. Measured 6.7× lower p99 TTFT than static batching under Poisson load. |
| 🎯 execlab | Quant / RL | Optimal-execution research lab — L2 replay simulator, TWAP / VWAP / Almgren–Chriss, and a PPO agent from scratch, with honest, ablation-tested findings |
| ⚡ nanolob | Systems / C++ | Low-latency C++20 limit order book & matching engine with Binance L2 replay and a market-making simulator |
| 🏎️ pitwall | Bayesian ML / Simulation | F1 race-strategy Monte Carlo optimizer. A hierarchical Bayesian tyre-degradation model that separates fuel burn from tyre wear — perfectly collinear within a stint — using a hand-rolled Gibbs sampler, feeding an ensemble-vectorized lap-by-lap race simulator and a reactive safety-car policy. Backtested on a held-out season: 2.47 places MAE, Brier 0.120 vs 0.250 climatology. The ablation contradicted my own hypothesis and is written up that way. |
| 🧩 Neuro-Symbolic RCA | ML / NLP | Hybrid neuro-symbolic approach to root-cause identification in customer-service conversations |
| 🧭 ADAPT | Quant | Regime-aware portfolio framework for Indian equities — market classification, multi-sleeve allocation, factor optimization vs NIFTY 50 TRI |
| 💳 Riskon | Fintech / ML | Dynamic credit-scoring algorithm |
🗳️ New — raftkv: a distributed, linearizable key–value store built on a from-scratch Raft implementation — fsync WAL, lease reads, exactly-once sessions, and dynamic sharding with live migration — all verified by a from-scratch linearizability checker running 1000 randomized fault schedules on every CI push. The bugs it caught (including two only CI found) are written up with root causes in
BUGS.md.⚡ Also — tinyserve: the inference-serving machinery behind systems like vLLM, implemented by hand — forward pass, paged memory manager, continuous-batching scheduler, and speculative decoding — with the static-vs-continuous batching benchmark documented end-to-end.
💡 Also start here →execlab's README walks through how a promising ML result turned out to be an artifact — and how a fair baseline exposed it. That mindset (breaking my own results before trusting them) is what I bring to everything.
⚡ "Build it, measure it, and try to break it — the market (and reality) will if you don't."