Skip to content
View pes1ug23am910's full-sized avatar
  • PESU
  • Bengaluru

Highlights

  • Pro

Block or report pes1ug23am910

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
pes1ug23am910/README.md

Hi, I'm Yash Verma πŸ‘‹

Final-year B.Tech β€” Computer Science (AI-ML) @ PES University, Bengaluru (2023–2027) Β· CGPA 8.96/10, merit scholarship every semester to date

I build reliability-focused systems in Rust, TypeScript, and Python, and applied-ML systems end to end β€” with real test suites, CI, and honest status reporting. Co-author + Best Presenter, ICCEE 2026; first-author NLP study in preparation. Seeking SDE / backend / systems and AI-ML internships and 2027 new-grad roles.


πŸ”§ Featured Projects

Raft consensus implemented from scratch in Rust β€” no consensus libraries of any kind β€” as a 3-node key-value store over TCP. Leader election, log replication with conflict-only repair, and the Figure-8 current-term commit rule, over durable storage with atomic hard-state swap and torn-tail recovery. The consensus core is I/O-free, so identical code runs against real sockets and against a virtual network.

Rust Cargo workspace TCP β€” deterministic virtual-time simulator asserting Election Safety / Log Matching / State Machine Safety after every step; failures replay from their seed

Adaptive retrieval-augmented generation: a zero-cost logit/entropy probe decides when to retrieve, RoBERTa-large-NLI scores passages, and hybrid FAISS-HNSW + BM25 retrieval runs over a 545k-passage index. 4-bit quantised Mistral-7B inference; claim-level provenance evaluation (FactScore-style).

Python Mistral-7B FAISS RoBERTa-NLI BM25 β€” pytest + ruff + mypy via GitHub Actions

Multi-agent tutoring system on Google ADK + Gemini 2.0 Flash: one orchestrator + 5 specialist agents + 3 validator loop-agents routing tutoring, study planning, adaptive quizzing, and progress analytics. 7-stage adaptive spaced repetition (Ebbinghaus curve) with JSON-based cross-session continuity.

Python Google ADK Gemini 2.0 Flash asyncio

Privacy-first local document processing β€” a typed Python core library, Typer CLI, and a token-authenticated loopback FastAPI service with Job-Object-contained workers. Structural and render validation before atomic, collision-safe publication; hash-locked dependency profiles and a reproducible-build gate. Also ships an MCP server so local agents can call it instead of hand-rolling conversion code.

Python FastAPI Typer pikepdf PDFium β€” 639-outcome test suite (636 passed, 3 documented platform skips at the last gated merge), ruff + mypy, release-gated

πŸ§ͺ Also building (private / local for now)

  • ASCEND β€” hybrid event-log/cached-projection desktop productivity RPG (Rust, Tauri 2, React, TypeScript): 11 replayable gameplay/economy projections rebuild from the append-only event log; core economy mutations pair cache and event writes in one SQLite transaction, with compensating transactional undo. 456 automated tests (327 Rust + 129 Vitest).
  • FocusMe β€” cross-platform focus-enforcement system in Rust: Windows (WFP) backend implemented and end-to-end tested, remaining backends in progress. SQLCipher-encrypted policy store; Axum backend with JWT auth and multi-device policy sync.
  • PromptGFM-Bio β€” prompt-conditioned graph model for rare-disease gene ranking (PyTorch Geometric, frozen PubMedBERT, FiLM conditioning); manuscript in preparation.
  • UI-Migration Capstone (Lead) β€” React Native β†’ Kotlin via a typed intermediate representation, with CodeT5+ (220M Γ—2) encoders; frozen-instrument exact-match reconstruction rose 17.8% β†’ 85.6% across model versions.

πŸ“„ Research & Publications

  • An Engineering-Oriented ML Decision Support System for Outcome Prediction in Dynamic Environments β€” ICCEE 2026 (19th Int'l Conference on Computer & Electrical Engineering, Brisbane) Β· IOS Press proceedings Β· Paper IC-1015 Β· Co-author, accepted & presented β€” and Best Presenter award for the presentation.
  • Why VLMs Fail on Indic Memes: A Failure Taxonomy of Cultural-Knowledge Gaps β€” first author Β· pre-registered (OSF, Apr 2026) Β· manuscript in preparation.

πŸ… Achievements

  • Best Presenter β€” ICCEE 2026, Brisbane (co-sponsored by Central Queensland University), for the online presentation of co-authored paper IC-1015
  • Selected β€” ISRO research internship, Space Applications Centre (SAC), Ahmedabad (SRTD, May 2026); could not join due to an academic-calendar conflict
  • Prof. MRD Scholarship β€” top 5% of CS (AI-ML), Sem 1 (SGPA 9.36) Β· Prof. CNR Scholarship β€” top 25%, Sems 2–5
  • Google Γ— Kaggle 5-Day AI Agents Intensive β€” multi-agent systems, tool use & orchestration (Google ADK), Nov 2025

πŸ› οΈ Tech Stack

Languages Python, Rust, TypeScript, JavaScript, Kotlin, C, SQL
ML / AI PyTorch, PyTorch Geometric, HuggingFace Transformers, BERT / PubMedBERT, FAISS, XGBoost, Google ADK, Scikit-learn
ML Concepts RAG, GNNs, NLI, transfer learning, model quantization, vector search, multi-agent orchestration, evaluation design (bootstrap CIs, pre-registration)
Systems & Desktop Electron, Tauri 2, Node.js, React, Win32, Axum, Flask, FastAPI, PostgreSQL, REST APIs
Reliability & Testing Event sourcing, atomic persistence & crash recovery, IPC security boundaries, Rust / Vitest / Node / pytest testing, GitHub Actions CI/CD
Tools Git, Docker, Linux, Weights & Biases, Gemini / OpenAI / Anthropic APIs

πŸ“¦ Earlier projects

Japanese Novel Translator Β· Kakuyomu Translator Β· Novel Scraper Toolkit Β· Study Tracker Β· Notion-Anki Sync Β· Habitify Dashboard Β· Python Utility Tools Β· Cab Aggregator

πŸ“Š GitHub Stats

GitHub Stats

Top Languages

πŸ“« Contact


"The best code is the code you actually use every day."

Pinned Loading

  1. micro-raft micro-raft Public

    Raft consensus implemented from scratch in Rust β€” 3-node key-value store over TCP, verified by a deterministic fault-injection simulator

    Rust

  2. LocalDocForge LocalDocForge Public

    Privacy-first local document processing: typed Python core library, CLI, and localhost API. Fully local - no cloud conversion, no telemetry.

    Python

  3. study-buddy-final study-buddy-final Public

    Multi-agent AI learning companion built with Google ADK and Gemini 2.0 Flash. Implements spaced repetition, adaptive quizzing, personalized tutoring, and intelligent progress tracking.

    Python 1