Skip to content
View Somansh1's full-sized avatar
3 followers · 3 following

Achievements

Achievement: Pull Sharkx2Achievement: Pair Extraordinairex3Achievement: QuickdrawAchievement: YOLO

Achievements

Achievement: Pull Sharkx2Achievement: Pair Extraordinairex3Achievement: QuickdrawAchievement: YOLO

Highlights

  • Pro

Block or report Somansh1

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. 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
Somansh1/README.md

Somansh Goel. Builds software that ships, then finds out how AI systems break. A star-trail sky over Himalayan foothills, observed from Dehradun.

Observer's note. I'm a final-year B.Tech CSE (AI & ML) student at UPES, Dehradun, graduating in 2027. I build web and desktop software end to end, and I spend the rest of my time finding out how AI systems fail.

Portfolio: somansh.in   LinkedIn   Email: somansh@somansh.in

A year of commits, charted

A polar star chart of the last 53 weeks of GitHub activity. Each day with commits is a star, the busiest days carry diffraction spikes, and the numbers beside it give the totals and streaks.

Catalogue. Six objects, each one shipped. Click a plate to open it.

SG 01, ChatLens. Chat analytics that never leave your browser. SG 02, SocialSpace. A table for two: chat, calls, a shared canvas and watch-together.

SG 03, Gem-VTuber. A desktop Live2D companion that talks through the Gemini Live API. SG 04, SRE Autopilot. An incident-response environment with seven microservices and a hidden root cause.

SG 05, AppD-Lite-Net. Apple leaf disease detection with a sub-million-parameter CNN. SG 06, Fouress Bath. The live catalogue site for a sanitaryware retailer.

Catalogue as text
Project What it is Built with
ChatLens Chat analytics that never leaves your browser. Parses WhatsApp, Instagram and Messenger exports and charts who texts more, when, and what. A Content-Security-Policy blocks every outgoing request, so the privacy claim is enforced, not promised. Source Next.js, TypeScript, hand-written SVG charts
SocialSpace A table for two. Chat, peer-to-peer calls, a shared canvas and watch-together between two people, over a small WebSocket relay and WebRTC. Source Next.js, WebSocket, WebRTC, Node.js
Gem-VTuber A desktop companion. A transparent always-on-top window renders a Live2D avatar that talks through the Gemini Live API; the model's tool calls drive its expressions and its lips follow the audio. Electron, PixiJS, Live2D, WebSocket
SRE Autopilot An RL-style environment for incident response. Seven microservices in a dependency graph, a hidden root cause, noisy telemetry, and an SLA-based reward for whichever agent is on call. Python, FastAPI, OpenEnv, Docker
AppD-Lite-Net Apple leaf disease detection with a CNN under one million parameters: 99.53% on a held-out test set and 98.80% in 5-fold cross-validation. Co-first author of the manuscript. Frontend TensorFlow/Keras, FastAPI, React
Fouress Bath The live catalogue site for a plumbing and sanitaryware retailer, with structured data and technical SEO done properly. Next.js, Tailwind CSS

Field notes: how assistants leak. A system prompt drawn as redaction bars while a scan line recovers the text beneath, with two 2025 security research entries.

Security research as text
  • Red-team audit of a production AI assistant (2025). Recovered the assistant's full system prompt through a chain of persona jailbreak, metadata probing and chunked extraction, then reported five findings to the vendor with a proposed hardened prompt design.
  • Prompt-extraction comparison across six consumer assistants (2025). Ran one family of extraction prompts against Gemini, ChatGPT, Grok, Perplexity, Meta AI and Claude: three returned instruction-like text, two refused, one returned nothing.

Logbook

Logbook. June to August 2025: AI/ML and Software Engineering Intern at Bharat Electronics Limited. Built a FAISS-backed retrieval pipeline on Llama 3.2 Vision over 100+ page internal documents, taking lookups for 50+ staff from hours to under 10 seconds. Scripted RHEL 9 environment setup, cutting per-machine provisioning from 3 hours to 15 minutes. Shipped a one-click installer so non-technical staff could set the system up on-prem. March 2026 to present: Secretary, IEEE Computer Society (UPES).

Experience as text

AI/ML & Software Engineering Intern, Bharat Electronics Limited · Jun to Aug 2025

  • Built a FAISS-backed retrieval pipeline on Llama 3.2 Vision over 100+ page internal documents, taking lookups for 50+ staff from hours to under 10 seconds.
  • Scripted RHEL 9 environment setup, cutting per-machine provisioning from 3 hours to 15 minutes.
  • Shipped a one-click installer so non-technical staff could set the system up on-prem.

Secretary, IEEE Computer Society (UPES) · Mar 2026 to present

Your move. This chess board is live: click a move below, press Create on the issue that opens, and the board redraws in about a minute.

Chess board with the move list, game 1, move 3, black to move, last move Nxe5

Black king on e8: Play Ke7
Black queen on d8: Play Qe7 Play Qf6 Play Qg5 Play Qh4
Black rook on a8: Play Rb8
Black bishop on f8: Play Ba3 Play Bb4 Play Bc5 Play Bd6 Play Be7
Black knight on c6: Play Na5 Play Nb4 Play Nb8 Play Nd4 Play Nxe5 Play Nce7
Black knight on g8: Play Nge7 Play Nf6 Play Nh6
Black pawns: Play a5 Play a6 Play b5 Play b6 Play d5 Play d6 Play f5 Play f6 Play g5 Play g6 Play h5 Play h6

No finished chess games yet. Game 1 is in progress.

Moves as text

Legal moves for black (32):

Recent moves:

    1. e4 by @Somansh1
  • 1... e5 by @Somansh1
    1. Nf3 by @Somansh1
  • 2... Nc6 by @Somansh1
    1. Nxe5 by @shreyamT

Instruments

Instruments. Languages: Python, TypeScript / JavaScript, Java, C / C++, SQL. Web: React, Next.js, Tailwind CSS, Node.js, WebSocket, WebRTC. AI / ML: RAG pipelines, FAISS, LLM fine-tuning, PyTorch, TensorFlow, scikit-learn. Systems: Linux (RHEL), Docker, Bash, Electron, Git.

Toolbox as text

Languages Python · TypeScript / JavaScript · Java · C / C++ · SQL
Web React · Next.js · Tailwind CSS · Node.js · WebSocket · WebRTC
AI / ML RAG pipelines · FAISS · LLM fine-tuning · PyTorch · TensorFlow · scikit-learn
Systems Linux (RHEL) · Docker · Bash · Electron · Git

Away from the keyboard: astrophotography, chess, and games. Every image here is hand-written SVG from scripts/build.py. Star chart redrawn nightly. No third-party widgets.

Pinned Loading

  1. SRE-RL-Env SRE-RL-Env Public

    Python

  2. aura-models aura-models Public

    Small fine-tuned models that drive the AURA Windows assistant

  3. BusinessER BusinessER Public

    Amazon ML Challange 2026

    Jupyter Notebook 1

  4. ChatLens ChatLens Public

    TypeScript 1

  5. SocialSpace SocialSpace Public

    TypeScript

  6. GemVTuber GemVTuber Public

    JavaScript