Engineering
i am currently a 2nd year b.tech cse student currently studying at manipal institute of technology in blr, india.
why engineering?? tbh idk. it's just one of those cool things ig? you build cool stuff thats cool and awesome.
{i seriously need to expand my vocabulary beyond the words cool and awesome}
now onto inspirations, one of my biggest inspirations in life was {and still is} apj abdul kalam. obviously his dedication and work ethic is a reason to be inspired but for me it's mostly his work in DRDO and ISRO.
he's prolly one of the main reasons {if not the only reason} why india has one of the best missile systems in the world.
and the other one is obviously tony stark. bro's so awesome man, how can anyone not be a fan?
ps: batman is better {because i am batman}.
Stack
- Languages:
- C, Python, Java, Bash
- DevOps:
- Docker, Linux (Distros: Mint, Fedora), Git and GitHub. Applications hosted: Immich, SearXNG
- Tools and Applications:
- Node.js, OpenCode, Ollama, Supabase, Expo.dev, Google Cloud.
- Libraries from my projects
- Flask, Matplotlib, NetworkX, NumPy, pandas, Pillow, PyTorch, scikit-learn, Shapely, TensorFlow, Hugging Face Transformers, Kaggle Datasets
Projects
Hub Router built
it's an application to place fire-alarm hubs and routes in simple closed, non-branching cable ring per hub on residential floor plans. built for the HackHive hackathon organized by Honeywell in 2026.
- ABOUT
- takes a residential floor plan (vector based) and decides how many fire-alarm control panels to place. where and which rooms each hub serves, and then finally routes one closed, non-branching cable ring per hub. cable runs either horizontally or vertically {no diagonals}, through doors only, never through walls or windows. a validator re-checks every rule and a renderer draws the blueprint overlay. a flask app demos it in the browser.
- HOW IT WAS BUILT
- built it with opencode, after researching the problem and the specs. we built it for the HackHive hackathon. it was disqualified at a checkpoint round because it wasn't compliant with the given specs (we had allowed diagonal placement of cable). nonetheless we went back and rebuilt it to the specs even though we didn't win and it prolly didn't matter.
- STACK
- Python, Shapely, NetworkX, NumPy, Matplotlib, Flask, Pillow
- GITHUB REPO
- github.com/ishan-bharath/HackHive
MIA v1 UNFINISHED
An AI companion trained on 30 example conversations. It ran end to end but the replies were repetitive.
- About
- The goal was a companion that talks like a friend and helps like a therapist. For the Elenxia Edutech hackathon I trained a small LSTM network, a model that reads text one word at a time and predicts what comes next, on 30 example conversations in a Google Colab notebook. The code runs from start to finish, but the replies came out generic and repetitive, so the project never became usable.
- What I did
- Wrote the training notebook: text tokenisation, the LSTM layers and the training loop.
- Stack
- Python, TensorFlow, scikit-learn, pandas, NumPy
- Links
- Repo
MIA v2 UNFINISHED
MIA v1 rebuilt on GPT-2 and fine-tuned on 150 conversations. The replies copied the training data, so I stopped.
- About
- After the hackathon I rebuilt the same companion on GPT-2, a small language model, fine-tuned on 150 example conversations with the Hugging Face Transformers library. The replies sounded more natural but stayed robotic, and the model often copied the training data instead of answering. In testing the replies scored 3 out of 10 for quality, and with this little training data I saw no clear path forward, so I stopped work and started redesigning the project from scratch.
- What I did
- Fine-tuned GPT-2 on 150 example conversations and tested the replies.
- Stack
- Python, PyTorch, Hugging Face Transformers, pandas
- Links
- Repo
APEX LIFE PERSONAL USE
- ABOUT
- a simple web app that acts a "life organizer" {its actually a plain to do app with AI integrations}. it's got sections for your daily to-do's {you can add more or ask the in-built AI chatbot to modify it accordingly} , your career goals, your health goals and a friendly AI chat interface that can remember your details, modify them accordingly and also modify your to-do etc. and all of this is private because you bring your own key, if needed you could also run it with help of local llms with an openai compatible end point.
- STACK
- HTML, CSS, JavaScript, Service Worker, localStorage, OpenRouter API
- Links
- Repo
Home server PERSONAL USE
a home server i run on a dell latitude laptop with linux mint.
- ABOUT
- a dell latitude laptop running linux mint that i use as a home server. it runs immich for photos and nextcloud for files (next cloud part ain't implemented yet).
i also run searxng, a private search engine, via docker on my desktop. - STACK
- linux mint, docker, immich, nextcloud, searxng
Hackathons
Elenxia
national level online hackathon. built MIA v1 for it. Got selected for top 10 but unfortunately couldn't complete on time. Idea for a AI therapist chat bot was then scrapped and then i started on MIA v2 but also hit a wall there due to enormous amount of data and compute required for re-training llms. eventually scrapped entire idea.
HackHive
2026participated and got selected in the top 30 teams out of 60+ teams that participated. unfortunately got disqualified during an intermediary checkpoint round, due to project specs not being compliant with given specs.
{nonetheless, we built a spec compliant version and it works according to required specs as given in the initial problem statement}