WhatsApp automation for Indian D2C brands, plus applied ML and embedded systems in Python. 2nd-year Electronics & Telecommunication at KJ Somaiya College of Engineering (CGPA 9.15/10).
I build and sell WhatsApp automation — order confirmations, delay alerts, keyword replies, broadcasts, AI replies with human handoff — and I ship fixes into projects I don't maintain. Currently taking on a small number of automation projects; details at dawnofgenx.github.io.
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.
I build and run a WhatsApp automation product for D2C brands — keyword rules, a visual flow builder, bulk broadcasts, AI replies with human handoff, and a shared team inbox.
- 81-second product demo — a recording of the live deployment, not a mockup
- 7 articles on dev.to — WhatsApp Business API, COD confirmation bots, broadcast at scale, API pricing in India, vendor comparison
- Vishwas — a WhatsApp-first verification platform, open source
Pricing is a two-week paid pilot on one flow, fixed price, no lock-in. I have not yet done paid client work, so I lead with the product, the code and the writing rather than a client list.
| 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. |
The WhatsApp automation product's source is private — it's client-facing work, not a portfolio repo. The demo video and the articles are the public window into it.
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.