Resume

B.Tech., Computer Science · Hyderabad, India

• Coursework: Data Structures & Algorithms, Operating Systems, Database Systems, Computer Networks, Machine Learning.
• IEEE Computer Society: Secretary, Student Branch Chapter; competed in IEEEXtreme 19.0, IEEE's global 24-hour contest (19,000+ participants).

Python, PyTorch, Polaris, Zenith, FastAPI, Docker

• Built a grounded RAG engine over PubMed literature with no third-party pre-trained models: from-scratch BPE tokenizer, InfoNCE dense bi-encoder, IVF/HNSW ANN indexes (no FAISS), cross-encoder reranker, and NLI verifier — every answer cited, verified, or refused, with calibrated abstention at 80.1% coverage and a 0.000 error rate.
• Trained the from-scratch NLI verifier 41.5% → 78.3% on SNLI dev (942K pairs, MLM pretraining over 1.15M sentences) by isolating three bottlenecks: tokenizer mismatch (+6.0 pts), data scale (+22.0 pts), and learning rate at identical architecture (+17.6 pts).
• Diagnosed a reranker overfitting collapse and restored R@5 0.029 → 0.983 via rank fusion; built HNSW reaching recall@10 0.996 at 0.262 ms, faster than exact search. Shipped v1.0 with 254 tests (95.7% coverage), seed-averaged ablations, and published negative results.

Python, PyTorch, FastAPI, Hydra, MLflow

• Built and published to PyPI a decoder-only transformer LM library written directly on PyTorch tensor primitives, implementing the modern LLM stack: causal self-attention, RoPE, RMSNorm, SwiGLU, KV-cache, weight tying. Matches the nanoGPT baseline at 2.08 bits/char, trained to parity in ~10 min on a laptop GPU.
• Implemented speculative decoding with KV-cache rollback (output provably identical to greedy, up to 3.7× fewer target forward passes) and from-scratch weight-only int8 quantization (4× smaller weights, output unchanged). Shipped v1.0 with 103-test CI and 22 tagged releases.

Python, PyTorch, FastAPI, Docker

• Architected the encoder-side complement to Zenith: an end-to-end NLP platform published to PyPI across 17 versioned releases — from-scratch BPE tokenizer, transformer encoder with no nn.Transformer, training, evaluation, and Dockerized FastAPI serving.
• Implemented BERT-style masked-language-model pretraining from scratch. In a controlled ablation against an identical randomly-initialized baseline, pretraining lifted first-epoch validation accuracy 73.6% → 81.0% and reached a higher best validation accuracy (86.4%) in fewer epochs.

TypeScript, Aiken, React

• Won the Masumi “Monetize AI Agents” track (Cardano; team of 4): forged data is traced to every agent that ingested it, directly or downstream, and revoked — with decontamination and auditing exposed as paid, hireable services across a 5-agent MIP-003 economy.
• Authored 3 Aiken (Plutus V3) validators enforcing quarantine in consensus logic; built SHA-256 sharding with Merkle non-membership proofs for provable deletion, plus an LLM pipeline with multi-provider failover (Groq → Gemini → deterministic). 82 tests (68 unit + 14 on-chain).

Hyderabad, India

• Built a multimodal audiobook pipeline (Whisper ASR → GPT-4 → ElevenLabs TTS) deployed to 500+ students; fine-tuned GPT-2 (124M) on ~500K tokens, cutting validation perplexity 28.5 → 24.2, served behind a FastAPI endpoint.
• Cut hallucination rate 42% → 27% on an internal 6-scenario evaluation via structured prompt engineering; added a GitHub Actions CI suite (90+ tests), cutting iteration time 45 → 18 min.