# Divanshu Sharma

> Machine Learning & AI Engineer. Founding Engineer @ Uniiq. Open to AI engineering & founding roles. Based in Mumbai, India.

Founding engineer who ships reliable AI products — currently leading the technical turnaround of Uniiq's AWS-deployed AI advising platform: 40+ critical vulnerabilities resolved, LLM sanitization and recoverable AI failures, LCP cut from 3.3s to 0.7s. Before that: privacy-preserving deep learning research at Yale and three years of quantitative research at WorldQuant BRAIN.

The depth behind it is public: a habit of rebuilding the modern AI stack from first principles — transformers, diffusion, RLHF, paged KV-caches, quantization, vector indexes — each published as a readable repository.

The through-line is reliable AI: structured outputs, prompt and response sanitization, failure recovery, rate limiting, consensus-fused ASR, and systems measured against explicit performance budgets.

Facts on this site come from his résumé, GitHub, repository READMEs and a Uniiq work summary; open-source counts are verified against the GitHub API. Nothing is inferred.

## Contact
- Email: divyanshu74.80@gmail.com
- Résumé (PDF): https://sdivyanshu90.github.io/sdivyanshu90/Divanshu_Resume.pdf
- CV (HTML): https://div90.vercel.app/cv
- GitHub: https://github.com/sdivyanshu90
- LinkedIn: https://www.linkedin.com/in/divsha22/
- Kaggle: https://www.kaggle.com/divanshu22
- LeetCode: https://leetcode.com/u/emotional_fool/
- CodeChef: https://www.codechef.com/users/psyduck90

## Experience
- Founding Engineer, Uniiq (May 2026 — present): Building Uniiq's AI-powered college-advising platform across React, Express, MongoDB, and Gemini, pairing conversational guidance with professional counselors for global university admissions.
- ML Researcher, Yale University (Oct 2025 — Mar 2026): Privacy-preserving deep learning: a from-scratch DL library built around compiler-centric multi-party computation. See Fig. 02.
- Research Consultant, WorldQuant BRAIN (Nov 2022 — Jul 2025): Quantitative research: high-frequency alpha signals in Python on WorldQuant's simulation stack.
- B.E. Computer Science, University of Mumbai (2021 — Jun 2025): CGPA 8/10. Coursework: ML, deep learning, DSA, probability & statistics.

## Selected work
- [Uniiq — AI Student-Advising Platform](https://div90.vercel.app/work/uniiq-platform): 40+ critical vulnerabilities — resolved in inherited legacy code, including prompt-injection and data-leak risks. Stack: React, TypeScript, Express, MongoDB, Gemini.
- [Privacy-Preserving Deep Learning via MPC](https://div90.vercel.app/work/mpc-deep-learning): 88.08% test accuracy — ChestMNIST, strided-convolution CNN, 5 epochs @ batch 512. Stack: Python, NumPy, MPC, Compiler design, Cryptography.
- [Indic ASR Optimization & Consensus Evaluation](https://div90.vercel.app/work/indic-asr): >48% absolute WER reduction — FLEURS Hindi benchmark vs. baseline Whisper-small. Stack: Whisper, Transformers, SpecAugment, Beam search, Librosa.
- [RenAIssance OCR — Historical Document Pipeline](https://div90.vercel.app/work/renaissance-ocr): Word accuracy 97.96% → 98.17% — full corpus, after LLM post-processing. Stack: PyTorch, CRAFT, CRNN, RoBERTa, OpenCV.
- [Paged KV-Cache Engine (vLLM-style)](https://div90.vercel.app/work/kv-cache-engine): Copy-on-write forking — ref-counted blocks shared across forked sequences. Stack: Python, PyTorch, Scheduling, Memory management.
- [Graph Data Explorer AI](https://div90.vercel.app/work/graph-data-explorer): 765 nodes / 877 edges — typed knowledge graph from 19 source tables. Stack: FastAPI, React Flow, NetworkX, SSE, Gemini 2.5 Flash.
- [Deep Learning for CMS Detector Physics](https://div90.vercel.app/work/ml4sci-cms): 80% AUC — electron-vs-photon classification, custom ResNet-15. Stack: PyTorch, PyTorch Geometric, TensorFlow, CNN, GCN.

## Open source (merged upstream PRs)
- [Mastra (mastra-ai/mastra)](https://github.com/mastra-ai/mastra/pulls?q=is%3Apr+author%3Asdivyanshu90+is%3Amerged): 11 merged — 29 PRs and 18 bug reports to the TypeScript agent framework (Aug–Sep 2026). 11 merged directly — codemod and CLI fixes for its migration tooling — and all 18 bugs he found and patched are fixed upstream, one merged with him as co-author. #58 of 466 contributors by commits.
- [EleutherAI lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness/pulls?q=is%3Apr+author%3Asdivyanshu90+is%3Amerged): 4 merged — 25 PRs to the standard LLM evaluation harness (Aug–Sep 2026): 4 merged — CLI argument parsing, request-cache robustness, and a restored benchmark task — and 21 in review.
- [Experiential (experientiallabs)](https://github.com/experientiallabs/experiential/pulls?q=is%3Apr+author%3Asdivyanshu90+is%3Amerged): 2 merged — 20 PRs (Sep 2026): 2 merged model-provider fixes — paginated Anthropic model discovery and Gemini page-token handling — and 18 in review.
- [Sequre (0xTCG)](https://github.com/0xTCG/sequre/pulls?q=is%3Apr+author%3Asdivyanshu90): proposal, unmerged — 4 PRs bringing the Yale work upstream — CNN layers and an MPC ChestMNIST training pipeline (#41 alone: +3.5k lines across 54 files). Not merged yet; #42 is open.
- [p5.js Web Editor (Processing Foundation)](https://github.com/processing/p5.js-web-editor/pulls?q=is%3Apr+author%3Asdivyanshu90+is%3Amerged): 4 merged — 10 PRs (2023–24), 4 merged: a keyboard shortcut for adding files, the Korean translation of keyboard shortcuts, and two UI fixes.
- [OpenCode (anomalyco/opencode)](https://github.com/anomalyco/opencode/pulls?q=is%3Apr+author%3Asdivyanshu90): proposal, unmerged — 22 fix PRs to the open-source AI coding agent (210k★), Aug–Sep 2026 — 21 open, awaiting maintainer review.
- [Also merged: vulnerablecode, Music Blocks, PyNN](https://github.com/pulls?q=is%3Apr+author%3Asdivyanshu90+is%3Amerged+-user%3Asdivyanshu90): 3 merged — A license-URL fix to aboutcode-org/vulnerablecode, a close-button fix to Sugar Labs' Music Blocks, and a docs-layout fix to NeuralEnsemble's PyNN.

## Skills
- Languages: Python (primary) · C++ · TypeScript / JavaScript · Rust · SQL · Bash · Java
- Modeling & training: PyTorch · TensorFlow · Keras · scikit-learn · Hugging Face Transformers · PyTorch Geometric · CNNs / GNNs · Whisper · DeBERTa / RoBERTa
- LLM systems: Gemini · structured AI outputs · RAG · SFT · LangChain / LangGraph · Model Context Protocol · prompt engineering · Qdrant · Chroma · Ollama
- Serving & MLOps: FastAPI · Docker · ONNX · GitHub Actions CI/CD · MLflow · Airflow · AWS (SageMaker, EC2) · GCP · Azure · Streamlit
- Data: NumPy · Pandas · Polars · CatBoost · PostgreSQL · MySQL · MongoDB · SQLite
- Web & product: React · Next.js · Node.js / Express · Django / Flask · Tailwind CSS · server-sent events · system design

## Education
- B.E. Computer Science, University of Mumbai, CGPA 8/10 (2021 — Jun 2025)

## Machine-readable
- [dossier.json](https://div90.vercel.app/dossier.json): the full typed dossier
- [MCP server](https://div90.vercel.app/api/mcp): streamable HTTP; tools search_dossier, get_case_study, list_systems, match_requirements (paste a job description), get_contact
- [From-scratch index](https://div90.vercel.app/systems): 47 AI-stack reimplementations