About

I'm K Satya Sai Nischal, a machine learning engineer focused on natural language processing. I build ML systems from first principles: not because frameworks are bad, but because the only way to genuinely understand a transformer is to implement one — the tokenizer, the attention mechanism, the training loop, the inference path — and then make it fast, correct, and usable by other people.

Engineering philosophy

My projects follow a simple rule: no black boxes in the core. If Polaris does masked language modeling, the masking strategy, the objective, and the encoder are code I wrote and can defend — PyTorch supplies only tensor primitives, autograd, and optimizers. Abstractions are earned: in both projects an abstraction is extracted only when two or more concrete implementations demand it, a rule enforced through architecture decision records.

I care about the unglamorous parts of ML engineering: reproducible training runs, honest benchmarks, evaluation harnesses that can't be gamed, APIs that fail loudly, and documentation that treats the reader as a peer. A library that trains a model but can't explain its own numbers is half-finished.

Why from scratch?

Re-implementation is my learning strategy and my credibility signal. Reading a paper tells you what a mechanism claims to do; implementing it tells you what it actually does, where it breaks, and what the paper left out. Every component in Polaris and Zenith exists because I needed to understand it deeply enough to make engineering decisions about it.

What I'm learning now

• Efficient inference: KV-cache management, speculative decoding, quantization.
• Retrieval-augmented generation and grounding for high-stakes domains.
• Distributed training and the systems side of scaling.
• Evaluation methodology — how to measure models without fooling yourself.

Where I'm headed

The grounded biomedical direction I set out toward is now real: Meridian v1.0 answers questions over PubMed abstracts with per-sentence citations, entailment verification, and calibrated abstention — every model trained in-house. Next: closing its measured domain-transfer gap (SciNLI adaptation), scaling the corpus to ~200K abstracts, and training the cited-answer generator at scale. Alongside that, maturing Polaris and Zenith into libraries other engineers reach for when they want transparent NLP infrastructure.