DIV-1 · Divanshu Sharma

Seven figures

Selected work, with measured results

Each figure is problem → approach → results, with a receipt for where the numbers come from. The shipped ledger below lists what he built end-to-end and owned.

Figures

  1. Fig. 01Uniiq — AI Student-Advising PlatformFounding engineering · Full-stack AI product · 202640+ critical vulnerabilitiesresolved in inherited legacy code, including prompt-injection and data-leak risks
  2. Fig. 02Privacy-Preserving Deep Learning via MPCML research · Yale University · 2025–2688.08% test accuracyChestMNIST, strided-convolution CNN, 5 epochs @ batch 512
  3. Fig. 03Indic ASR Optimization & Consensus EvaluationSpeech · Evaluation research · 2025>48% absolute WER reductionFLEURS Hindi benchmark vs. baseline Whisper-small
  4. Fig. 04RenAIssance OCR — Historical Document PipelineComputer vision · NLP · 2026Word accuracy 97.96% → 98.17%full corpus, after LLM post-processing
  5. Fig. 05Paged KV-Cache Engine (vLLM-style)LLM inference systems · 2026Copy-on-write forkingref-counted blocks shared across forked sequences
  6. Fig. 06Graph Data Explorer AIApplied LLM product · 2026765 nodes / 877 edgestyped knowledge graph from 19 source tables
  7. Fig. 07Deep Learning for CMS Detector PhysicsScientific ML · ML4SCI / CERN CMS · 202480% AUCelectron-vs-photon classification, custom ResNet-15

Shipped ledger

Built end-to-end and shipped — what he owned, not what he studied.

  1. Uniiqlive productFounding Engineer — full-stack AI product

    Led the turnaround of an AWS-deployed AI onboarding platform: resolved 40+ critical vulnerabilities, added LLM sanitization and resilient conversational intake, expanded student and admin workflows, and raised web performance from 73 to 93.

  2. DIV-1 (this portfolio)0→1Designed, built & shipped solo

    An AI-native generative-UI console with a deterministic fallback and a server-side prompt-injection firewall — the interface you're using now.

  3. Graph Data Explorer AIfull-stack AIBuilt end-to-end

    Turned 21,393 SAP records into a guarded natural-language knowledge-graph query engine — FastAPI + SSE + React Flow, a two-stage LLM pipeline, three guardrail layers.

  4. University Admission MCPagentsBuilt end-to-end

    Agentic admissions advisor: Playwright scraping → SQLite → MCP server → tool-calling client, with 39 automated tests and strict typechecking.

  5. RenAIssance OCRapplied MLResearch + full pipeline

    Three-stage historical-OCR system (CRAFT → SE-ResNet-BiLSTM CRNN → RoBERTa) reaching 98.17% full-corpus word accuracy on 16th-c. Spanish print.

  6. GSoC ExplorerproductBuilt end-to-end

    Next.js app over five years of Google Summer of Code data, backed by a SQLite-cached archive API with analytics.