K Satya Sai Nischal
Machine Learning Engineer — NLP, retrieval & efficient inference
Building machine learning systems from first principles. I build open-source NLP systems end to end — tokenizers, transformer encoders and decoders, retrieval and grounding, training engines, and serving stacks — on raw PyTorch tensor primitives, and ship them as real libraries on PyPI.
- packages on PyPI
- 4packages on PyPImeridian-rag, polaris-nlp, zenith-nlp, komichi
- pre-trained models used
- 0pre-trained models usedEvery ML component in Meridian is trained in-house
- bits/char, nanoGPT parity
- 2.08bits/char, nanoGPT parityZenith's 10.7M-parameter from-scratch decoder
- test coverage
- 95.7%test coverage254 tests across Meridian, mypy --strict
01Selected work
Open-source ML systems, built from first principles
Each project is a documented case study — problem, architecture, design decisions, measured benchmarks, and honest negative results.
01Grounded RAG
Stable · v1.0Meridian
A from-scratch grounded RAG engine over biomedical literature (PubMed). Custom BPE tokenizer, dense retriever, HNSW/IVF ANN, cross-encoder reranker, and NLI faithfulness verifier — no third-party pre-trained models. Every answer is cited, verified, or refused.
System pipeline
- BM25
- Dense retrieval
- HNSW / IVF
- Cross-encoder
- Cited generation
- NLI verify
- Abstain
Measured results
- R@5 after reranker rank fusion
- 0.983
- HNSW recall@10 at 0.26 ms
- 0.996
- NLI verifier, SNLI dev
- 78.3%
- R@5 after reranker rank fusion
- 0.983R@5 after reranker rank fusion
- HNSW recall@10 at 0.26 ms
- 0.996HNSW recall@10 at 0.26 ms
- NLI verifier, SNLI dev
- 78.3%NLI verifier, SNLI dev
pip install meridian-rag02Encoder NLP
Stable · v1.1Polaris
A production-inspired NLP engineering platform for understanding text. The complete encoder-side stack — data, tokenization, collation, models, training, evaluation, deployment — from scratch on PyTorch tensors.
- IMDB ceiling, four-lever study
- ~86%
- epoch-1 val with MLM (vs 0.74)
- 0.81
- versioned PyPI releases
- 17
03Decoder LLMs
Stable · v1.0Zenith
A from-scratch generative NLP library — decoder-only language models and text generation. Llama-style internals (RoPE, RMSNorm, SwiGLU), KV-cached generation, speculative decoding, LoRA, instruction tuning. Matches the nanoGPT baseline.
- bits/char, nanoGPT parity
- 2.08
- fewer target passes, spec. decoding
- 3.7×
- smaller weights with int8
- 4×
04ML tooling
Stable · v2.0Michi
道 — a local-first ML workbench. Independent CLI tools that profile datasets, evaluate models, and benchmark them with real statistical rigor. Automates the implementation, never the judgement.
- independent CLI verbs
- 18
- leakage overstatement caught
- 26 pts
- network calls, fully local
- 0
02Toolkit
The stack I build with
Most of the ML below I've implemented myself rather than imported — which is what lets me debug it, measure it, and make it fast.
- Machine Learning
- PyTorchTransformers (from scratch)MLM / CLM pretrainingLoRAMixed precision (AMP)DDP
- NLP & Retrieval
- RAGBM25Bi- / cross-encodersHNSW / IVFInfoNCENLIBPE tokenizationEvaluation & ablations
- Inference
- KV-cacheSpeculative decodingint8 quantizationRoPESwiGLU / RMSNorm
- MLOps & Infra
- FastAPIDockerGitHub ActionsMLflowHydrapytest / mypyPyPI packagingAWSLinuxPostgreSQLRedis
- Languages
- PythonC++TypeScriptSQLBash
03Now
What I'm focused on
- Building
- Michi v2.0 just shipped. Next: SciNLI domain adaptation for Meridian's verifier and the 200K-abstract corpus.
- Researching
- Faithfulness verification, calibrated abstention, and efficient inference for high-stakes domains.
- Reading
- Hoffmann et al. on compute-optimal scaling; Leviathan et al. on speculative decoding.
- Recognition
- 1st place, IndiaCodex '26 — Masumi “Monetize AI Agents” track, with ANTIDOTE.
04Writing
Latest engineering notes
Write-ups from real project decisions and measured results.
The Cross-Validation Bug Worth 26 Points of Fake Accuracy
Michi warned users about group leakage while its own benchmark command ignored groups entirely. How entity leakage inflates scores, and why it had to be fixed before the API freeze.
When Your Reranker Is Worse Than Random
Meridian's cross-encoder pushed Recall@5 to 0.029 — below random ordering. Diagnosing the collapse, and why fallible pipeline stages must degrade gracefully.
Greedy-Exact Speculative Decoding
How Zenith uses a small draft model to cut target-model forward passes by 3×+ while producing output identical to greedy decoding.
Weight Tying in Language Models
Sharing the input embedding and output projection — why it works, when it stops working, and how it interacts with logit scale.
Open to ML engineering roles
Building something where the ML has to be right?
I'm looking for ML engineering roles, research collaborations, and open-source work. Email is the fastest way to reach me.