I'm a 2026 CSE graduate from IIT Guwahati. I work across backend systems and applied AI, and I'm most interested in problems where reliability and performance actually matter — storage internals, distributed systems, LLM infrastructure, the layers underneath the products people use.
const sai = {
role: "Software Engineer (2026 grad)",
school: "IIT Guwahati, Computer Science",
focus: ["Storage systems", "LLM evaluation", "Backend"],
languages: ["Rust", "C++", "Python", "Java"],
currently: "Building Sastran — a crash-safe storage engine",
competitive: "Codeforces Expert (max 1803)",
writing_at: "tensen.dev",
open_to: "Full-time SDE / Backend / Infrastructure roles",
based_in: "India",
availability: "Immediate"
};The fastest way to know what I work on is to look at what I build — scroll down.
-9B59B6?style=for-the-badge&logo=neo4j&logoColor=white&labelColor=0D1117){kind=icon}
|
Crash-safe storage engine in Rust Unified key-value and vector storage. LSM-tree for KV, HNSW for ANN search, one durability story across both.
|
Full-stack RAG document assistant Production-style RAG system supporting semantic Q&A across PDFs, DOCX, HTML, Markdown.
|
|
B.Tech thesis · Prof. Amit Awekar Cost-aware routing system that dynamically selects between small and large language models per query.
|
NLP / Deep Learning course project Fine-tuned DeBERTa-v3 for sarcasm detection with custom affective-feature fusion.
|
|
Track 01 hackathon · CPU-fast ranking with LLM coherence Ranks 100,000 candidates against a Senior ML/AI JD via structural signals + LLM-grounded coherence checking. 5-stage pipeline: relevance floor → honeypot gate → JD disqualifier → calibrated scorer × multipliers.
|
A few other things I've built
All public repos at my repositories page → |
A few principles I've actually learned from building these — not borrowed from somewhere.
Measure before optimizing. The 28× bloom filter speedup in Sastran came from writing the benchmark in Criterion before writing the optimization. The 41% routing cost reduction came from labeling 1,500 queries with GPU oracle data before training a router. In Redrob I dropped several intuitive signals — keyword density, duplicate-description detection, education date-ordering — to zero after measuring they had no ranking value across 100K candidates. I learned this the hard way enough times that it's now reflex.
Learn by building. Rust at depth came from building Sastran, not from a course. HNSW came from reading the Malkov–Yashunin paper and implementing it, not from a library wrapper. The fastest path to understanding something is usually to ship it.
Comfortable being wrong. From competitive programming, my code gets judged objectively against test cases every weekend. Disagreement and error feel like information, not threat.
Honest about scope. Sastran is a single-node engine, not a distributed system. DocuMind is production-style, not battle-tested in production. I'd rather scope something small and finish it well than wave my hands at something I haven't built.
|
|
|
I've competed on Codeforces since 2022. It's where I first learned to think about problems systematically and to take being wrong as information.
|
|
📈 Contribution activity graph
Reading shapes how I work as much as building does. A short list of what's actually on my shelf — split between systems papers that inform my engineering and the broader books that shape how I think.
| Paper | Why it matters to me |
|---|---|
| The Log-Structured Merge-Tree — O'Neil et al. | The foundation behind LevelDB, RocksDB, and Sastran's KV layer |
| Efficient ANN search using HNSW — Malkov & Yashunin | What Sastran's vector index implements |
| Dynamo: Amazon's Highly Available Key-value Store | For thinking about distribution, replication, and eventual consistency |
| The Google File System — Ghemawat et al. | Still the cleanest introduction to large-scale block storage |
| Book | Why it matters to me |
|---|---|
| Designing Data-Intensive Applications — Kleppmann | The textbook I wish every backend engineer read |
| Database Internals — Petrov | Dense but excellent on storage engines specifically |
| Book | Why it matters to me |
|---|---|
| Thinking, Fast and Slow — Kahneman | The book that changed how I treat my own intuitions |
| Deep Work — Cal Newport | Why I protect long uninterrupted blocks for hard problems |
| Atomic Habits — James Clear | Systems over goals — the framing that actually changed my daily work |
| Make It Stick — Brown, Roediger, McDaniel | Evidence-based learning, written for technical readers |
| Clear Thinking — Shane Parrish | On catching the moments where defaults make decisions for you |
| Man's Search for Meaning — Viktor Frankl | The book I return to when work feels heavy |
| Meditations — Marcus Aurelius | Short, ancient, still the best operating manual I've found for a noisy mind |
I write occasionally at tensen.dev — mostly notes on systems, what I'm learning, and the occasional deep-dive into something I built.
I'm open to full-time SDE, backend, and infrastructure roles, including AI / LLM platform work. Most interested in companies where the engineering bar is high and the work touches reliability, performance, or data at meaningful scale.
📍 Available immediately · Bengaluru / Hyderabad / Remote