Applied ML & embedded systems. 2nd-year Electronics & Telecommunication at KJ Somaiya College of Engineering (CGPA 9.15/10). I build verification and telemetry tools in Python, and I ship fixes into projects I don't maintain.
I work at the seam between applied ML and embedded: models that have to run on a CPU-only laptop, and protocols decoded off a serial line.
| Vishwas | ~34K LoC · 445 tests | Zero-cloud WhatsApp verification platform. A user messages a number; URLs, executables/APKs, documents (OCR) and images/audio get checked and answered with a plain-language verdict plus a confidence band, in 7 languages. Routing is deterministic — an LLM only narrates finished verdicts, behind a prompt-injection guard. |
| UARTScope | ~13.9K LoC · 160 tests | "Wireshark for microcontrollers." Bidirectional UART/I²C/SPI/CAN/Modbus decoders, CAN from vendor DBC files, live charts, threshold alerts, session record/replay with golden-baseline diffing, plugin marketplace, MQTT bridge. |
| CiteSure | ~9.7K LoC · 288 tests | Local citation verifier (MCP server + CLI). Checks whether a cited source actually supports the claim, not just whether the link is alive. Unverifiable claims return "can't verify" instead of a hallucinated verdict. |
| blind-mcp | ~1K LoC · 16 tests | MCP server for company research on Blind, extended from an upstream MIT scaffold. |
Everything above runs with no cloud dependency and no API keys. CiteSure's fast
path is fully offline with --no-nli; its NLI tier is opt-in and lazy-downloads a
local model on first use.
3 merged PRs into PyPA-ecosystem tooling:
- tox-dev/filelock#752 —
fix(read-write): honor instance defaults - tox-dev/filelock#751 —
fix(marker): preserve live owners - pypa/pipx —
fix(upgrade): don't claim "latest version" for local-path installs
9 open and under review: pytest, pytest-xdist, pandas, Hugging Face diffusers (×2), MCP reference servers, LMCache, Axl.
Also submitted fixes at sentry-python, setuptools, twine and towncrier.
The common thread: read the codebase, write the failing test first, make the smallest change that fixes the actual bug — not my symptom of it.
- Languages — Python, C, TypeScript
- ML/AI — PyTorch, Hugging Face Transformers, LLM fine-tuning & evaluation, retrieval, MCP, OCR, deepfake detection
- Backend — FastAPI, REST, WebSockets, pytest, GitHub Actions CI
- Embedded — UART, I²C, SPI, CAN (DBC), Modbus, MQTT, ESP32, Raspberry Pi, Arduino, Betaflight
I keep a public devlog — one dated entry a day, each curating real arXiv cs.LG papers, so the learning is verifiable rather than claimed.
- Email — pkansara0412@gmail.com
- LinkedIn — linkedin.com/in/priyansh-kansara-a35826368
- Portfolio — https://dawnofgenx.github.io
- dev.to — https://dev.to/dawnofgenx
Open to ML-engineer internships in applied ML, LLM inference / edge deployment, or Physical AI — available for 3-week windows from mid-December 2026.