Skip to content
View DevGurav's full-sized avatar
8 followers · 48 following

Block or report DevGurav

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.

Content in all repositories owned by your account will be closed.
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
DevGurav/README.md

Hi, I'm Devendra Gurav

Typing SVG

About Me

  • 🎓 Final Year B.E. in Artificial Intelligence & Data Science — VCET, Mumbai University
  • 🤖 Building intelligent systems at the intersection of ML, backend APIs, and real-time data pipelines
  • 🛡️ Shipped Fall Guardian v3 (wearable AI fall prediction), hark (deterministic record & replay for AI agents, in Go), UrbanHeat AI (Mumbai heat-island digital twin) ZK-PoC (verifiable browser compute — 3 npm packages, live demo) and unsaid (a checker for characters TTS models silently refuse to say) end-to-end
  • 👁️ Shipped Drishti, a fully offline vision assistant for blind users in Marathi, Hindi and English — five modes, zero cloud calls, and a negative fine-tuning result published rather than buried
  • 🔍 Actively seeking SDE / ML internship & placement opportunities for 2026
  • 📫 Reach me at dev.gurav011@gmail.com

Tech Stack

Languages

Python Java Go C++ JavaScript TypeScript Dart SQL Solidity

Frameworks & Libraries

FastAPI Next.js React Three.js Flutter PyTorch scikit-learn XGBoost LangGraph ONNX Pandas NumPy Netty

Databases & Infrastructure

PostgreSQL Redis Supabase Docker GitHub Actions Git Jupyter

Featured Projects

✅ completed  ·  🚧 in progress — open Highlights on any card for the technical deep-dive

🛡️ Fall Guardian v3  

Wrist-worn AI that predicts falls for elderly users before impact — an edge model fires in <80ms, a cloud model confirms and grades severity, and the caregiver's phone gets a live alert.

Highlights
  • 96.5% recall · 0.7% false-positive rate on daily activity · ~300ms pre-impact lead time
  • Edge ConvLSTM-tiny (INT8, ~46 KB) + cloud Transformer (ONNX, 5-fold cross-validated)
  • Live SSE + FCM alerts, emergency SOS, per-user active learning from cancelled alarms

📼 hark  

An agent runtime that records an AI agent at a boundary it cannot bypass, replays the run deterministically, and proves the replay is real — kernel-enforced containment and a transparency-log anchor in one artifact.

Highlights
  • 25/25 runs replay-equal across 5 agent shapes — streaming, retry-after-503, repeats and MCP included
  • A Merkle Mountain Range proves one event happened in 448 bytes, without disclosing the rest of the run
  • Mediation costs +0.18 ms at p50; a replay that skips a 0.9 s model call finishes in 145 ms
  • A prompt-injected agent's exfiltration is denied by the kernel — and the key it leaked was a placeholder

🔐 ZK-PoC  

Asks whether consented, cryptographically-verified spare browser compute could fund the web instead of ads — and measures the answer instead of asserting it. Three npm packages, a live WebGPU demo, and a correction to a published IEEE paper.

 

Highlights
  • Found the canonical IEEE TDSC 2024 economic baseline overstated by ~87× — every other figure in the paper reproduces exactly, so the correction strengthens their conclusion
  • Derived the never-published break-even resource share: σ* = 6.7% against a 5% design ceiling — off by 1.3×, not an order of magnitude
  • 3 npm packages published; in-browser Groth16 proofs over a quantized compute circuit
  • Every economic and security claim is measured and reported, including unfavourable ones

🌆 UrbanHeat AI — Mumbai  

Maps and explains Mumbai's urban heat islands from satellite data, ranks the worst-hit wards, simulates the cooling effect of interventions, and answers planner questions through an AI copilot.

 

Highlights
  • Predicts an unseen ward's surface temperature to 1.10 °C (spatial R² 0.893) from Landsat LST, SHAP-explained per cell
  • Digital twin: greening cools 7,410 cells by −0.65 °C on average — and the 482 cells it warms are reported, not hidden
  • 4 LangGraph agents (planning, digital twin, monitoring, RAG copilot) deployed and live
  • Built end-to-end on free tiers and open data — no paid services anywhere

🗄️ Mnemo — Redis Built From Scratch  

A Redis-compatible in-memory store hand-built in Java 21 — custom hash table, skip list, and LRU/LFU eviction under the real RESP2 protocol, so redis-cli connects with zero changes.

Java 21 Netty RESP2 Docker

View Repository

Highlights
  • 39 commands · 126 tests green · 14 Architecture Decision Records
  • Incremental rehashing, span-augmented skip list, AOF crash recovery
  • CRC-16 sharding across N lock-free executor threads via Netty

👁️ Drishti  

A fully offline assistant for blind users that reads medicine strips, identifies rupee notes, reads Devanagari text, and describes surroundings aloud in Marathi, Hindi, and English — no internet, no cloud upload.

Highlights
  • All five modes run end-to-end on real models — read, medicine, currency, scene and visual Q&A
  • Currency classifier at 98.27% over 7 note classes, from a 6.2 MB checkpoint
  • Medicine mode is guardrailed against India's NLEM 2022 list — the VLM never guesses a drug name
  • Prompt engineering lifted VizWiz 0.308 → 0.533 with no training, so the LoRA was held to 0.533. It ran twice, came back statistically indistinguishable, and ships as a reported negative result — the prompt shipped, the adapter did not
  • The offline claim is measured, not asserted: the full task set run in aeroplane mode, spoken output verified by ear in both languages
  • 255 tests run with no GPU and no model weights; limits stated rather than softened — it does not claim a blind user can operate it

🔢 unsaid  

Text-to-speech models silently drop characters they cannot encode — no error, no warning, nothing in the logs, so the text you logged is correct and the audio your user heard is not. unsaid catches those characters in your test suite. The measurement behind it says the problem is the norm, not one model's quirk.

Python Transformers SentencePiece pytest GitHub Actions

View Repository Coverage Table

Highlights
  • Measured 30 TTS front-ends: only 4 of the 28 analysable ones can say all ten digits. 18 support some digits and not others — the dangerous shape, because numbers come out wrong rather than empty
  • The English model mms-tts-eng cannot say 7, 8 or 9; mms-tts-deu is missing only 9, so nine numbers in ten are perfect and nobody investigates the tenth
  • Came from a real incident: a Marathi voice has no 3, 5 or 8, so a Rs 500 note was announced as "00 rupees" for twelve days, past a 237-test suite that only ever asserted on the text
  • Each backend mirrors its front-end's normalisation from upstream source, then contradicts itself against upstream's own output — a disagreement is raised as a bug in the tool, never as a finding about your text
  • Refuses to guess: a model that romanises or phonemises before the vocabulary is consulted is reported as not analysable rather than given a confident, meaningless answer

Also: 3D Book Reader — open any PDF and read it on a realistic 3D book in the browser, with a dyslexic-friendly reflow mode and text-to-speech, 100% client-side (live demo)

GitHub Stats

Every graphic below is rendered daily by GitHub Actions and committed to this repo, so the page loads only files GitHub itself serves — there is no third-party service that can rate-limit, pause, or go down and leave a broken image here.

GitHub contribution snake (dark) GitHub contribution snake (light)

DevGurav's GitHub stats   DevGurav's most used languages

profile views

Pinned Loading

  1. fall-detect-system fall-detect-system Public

    Fall Guardian: An end-to-end, event-driven fall detection system. Features a FastAPI microservice backend, Redis Pub/Sub for live Server-Sent Events (SSE), PostgreSQL with Row-Level Security, machi…

    Jupyter Notebook

  2. hark hark Public

    Deterministic record and replay for AI agents — with a proof the replay is real. Kernel-enforced containment, tamper-evident audit, forkable runs. Linux, Go.

    Go

  3. zkpoc zkpoc Public

    Consent-governed browser compute, measured not assumed. Corrects a cryptojacking economics paper's ~87× error, publishes a break-even surface vs. ads, and replaces wasted CAPTCHA-style hash puzzles…

    JavaScript

  4. Drishti Drishti Public

    Offline AI vision assistant for blind users in India — reads medicine strips, identifies currency, and describes scenes aloud in Marathi, Hindi, and English. No internet, no cloud upload.

    Jupyter Notebook

  5. urbanheat-mumbai urbanheat-mumbai Public

    AI-powered urban heat island platform for Mumbai — predicts surface temperature per 200m grid cell from satellite data, explains drivers with SHAP, simulates cooling interventions, and answers plan…

    Python

  6. mnemo-redis-compatible-server mnemo-redis-compatible-server Public

    Mnemo is a high-performance, single-node, Redis-compatible in-memory data store built from scratch in Java 21. It features a custom lock-free data plane with an intrusive object-pooled Dict to mini…

    Java