About
Who I Am
My Journey
I'm Vansh, a third-year CSE student specializing in AI/ML at PES University. Most of my work now sits close to the machine: performance and correctness inside a production LLM fine-tuning runtime, a storage engine I'm writing from scratch in C++, and services that have to hold up against a real timeout. The part I care about is the one that usually gets skipped, which is proving a change did what you said it did. Right now I'm looking for backend and ML systems work where the problems are hard and the feedback is real.
Timeline
2022 — 2024
2024 — 2028
Aug 2025 — Jul 2026
Jun 2026 — Jul 2026
Cambridge International School
High School Diploma
Sciences & Mathematics through Grade XII, in Amritsar.
PES University
B.Tech CSE (AI/ML)
Computer Science with an AI/ML specialization, in Bengaluru.
AIESEC in Bengaluru
Senior Manager, oGT
Promoted within 6 months; led every team in the Outgoing Global Talent department.
PESU Research Foundation
Product Intern
Fine-tuned Qwen 2.5-VL 32B on an 865-prescription dataset (700/85/85 train/test/val split) to replace third-party OCR APIs, reaching 95% exact-match accuracy on medicine name, dosage and frequency on the held-out test set. Built Sanjeevani AI's medicine-lookup layer: fuzzy matching over a 250,000+ drug dataset, SQLite caching and an OpenFDA fallback, shipping multilingual TTS across 22 Indian languages.
2022 — 2024
Cambridge International School
High School Diploma
Sciences & Mathematics through Grade XII, in Amritsar.
2024 — 2028
PES University
B.Tech CSE (AI/ML)
Computer Science with an AI/ML specialization, in Bengaluru.
Aug 2025 — Jul 2026
AIESEC in Bengaluru
Senior Manager, oGT
Promoted within 6 months; led every team in the Outgoing Global Talent department.
Jun 2026 — Jul 2026
PESU Research Foundation
Product Intern
Fine-tuned Qwen 2.5-VL 32B on an 865-prescription dataset (700/85/85 train/test/val split) to replace third-party OCR APIs, reaching 95% exact-match accuracy on medicine name, dosage and frequency on the held-out test set. Built Sanjeevani AI's medicine-lookup layer: fuzzy matching over a 250,000+ drug dataset, SQLite caching and an OpenFDA fallback, shipping multilingual TTS across 22 Indian languages.
Technical Arsenal
Languages
Backend & Systems
ML & Data
Testing & Tooling
Frontend
Contribution
Open Source
Soup
Open-source LLM fine-tuning CLI
Sep 2026 – Present
Contributor, 6 merged pull requests
- Retimed large-layer prefetch in the layer-streaming runtime (#1178, #1258). Moved the embedding load to the backward tail and the LM head's load to layer 0. Median per-step wait fell from 15.8ms to 1.4ms on the disk tier and from 12.0ms to 4.5ms on the RAM tier.
- Verified gradients and final adapter weights stayed bit-identical across SFT, ORPO, SimPO, DPO and KTO, pinned by GPU tests that fail when the stream-ordering guard is removed.
- Built an interleaved A/B benchmark harness with an A/A control arm and provenance-stamped JSON output (GPU, driver, PCIe link, full argv, source commit) so the published figures are reproducible.
- Smaller PRs: removed orphaned BitNet helpers; collapsed nine duplicated control-strip tables onto a shared helper behind an AST-walking guard; repinned a published benchmark notebook and re-ran it on real hardware after finding its forecast rested on an assumed clock rate.
Work
Select Projects
minidb
LSM-tree key-value storage engine, written from scratch
A persistent key-value engine built with no storage libraries: a write-ahead log of CRC32-checked records, a hand-written skip-list memtable, and flushes to immutable on-disk SSTables, with reads and range scans merging the memtable and all flushed files newest-to-oldest. Crash recovery is derived from an explicit durability contract: a truncated tail is forgiven because the write was never acknowledged, while mid-file corruption fails the open rather than silently discarding writes the caller was told had succeeded. Tested with a differential oracle against a std::map reference model and a SIGKILL crash-injection harness asserting every acknowledged write survives. Compaction and benchmarks against SQLite are not built yet.
Codex Reviewer
Repository-aware automated PR review
A deployed GitHub App that ingests pull-request webhooks and posts findings as inline review comments. Changed files are parsed with tree-sitter into ASTs across JavaScript, TypeScript and Python, so the model receives function and class boundaries rather than raw diff text. A convention-profiling CLI mines the repository's existing patterns and flags only the deviations a PR introduces, applying and testing each generated fix in an isolated copy first. Webhook ingestion is decoupled from analysis by a BullMQ/Redis queue and a separate worker, keeping acknowledgement inside GitHub's 10-second timeout regardless of PR size, and model output is Zod-validated so malformed responses retry instead of posting.
Sanjeevani
Most people can't read their own prescriptions. They're dense, technical, and often in languages patients don't speak. Sanjeevani lets you scan a prescription and get a plain-language audio explanation — in 22 languages, in real time.
ClearTriage
Emergency rooms make life-or-death prioritization calls under pressure, often with no systematic support. ClearTriage is a real-time triage system that ranks patients by severity and shows why — using Explainable AI so doctors understand the reasoning, not just the output.
- Random Forest ESI classifier with SHAP explainability surfacing a per-prediction rationale.
- Validated with a 50-case human-vs-AI clinical audit; load-tested at 1,000 concurrent users with 0% errors and +82ms added latency.
Contact
Let's Connect
Start a Conversation
I'm currently open to new opportunities. Whether you have a question, a project idea, or just want to say hi, my inbox is always open.
vansh@vansharma.tech