Not a notebook. A system.
Three AI systems built with a team, each taken from raw data through to a service that runs. Retrieval, chatbots and computer vision in Python — AI & ML graduate in Thrissur, Kerala.
Ask a bad question.
Get the right answer.
Keyword and semantic search run in parallel, then the two rankings fuse. A query lands even when nobody knows the official name of the thing they're looking for.
Three systems deep, and just getting started.
I'm Goutham Krishna K S, a 2026 B.Tech graduate in Artificial Intelligence and Machine Learning from Vidya Academy of Science & Technology, Thrissur.
Everything I've built so far has been built with a team, and the part I keep gravitating toward is the layer underneath — the retrieval, the data pipeline, the thing that decides which answer actually comes back. On SevanaGPT that meant hybrid search over 4,600+ welfare schemes: keyword and 1024-dimension vectors running in parallel, fused with reciprocal rank fusion, so a query lands even when nobody knows the scheme's official name.
Alongside that I run a small self-hosted Docker setup at home — somewhere between six and twelve containers depending on what I'm curious about that month. It started as a way to stop breaking my own machine, and turned into the reason Compose files, volumes and networking aren't theory to me.
When I'm not on a personal project I'm usually out somewhere. Both, it turns out, are good for figuring out what a thing was actually supposed to do.
- Name
- Goutham Krishna K S
- Degree
- B.Tech — Artificial Intelligence & Machine LearningClass of 2026
- Institute
- Vidya Academy of Science & TechnologyAPJ Abdul Kalam Technological University
- Based in
- Thrissur, KeralaIndia · IST (UTC+5:30)
- Focus
- Retrieval systems · applied NLPPython · FastAPI · Docker
- Homelab
- 6–12 self-hosted containersDocker Compose · Linux
What I reach for.
Grouped by where it sits in a system rather than by how confident I feel about it. Everything here has been used in something on this page.
What I've built.
Team projects, each taken from raw data through to something that runs. Source is public — open it and read the code.
SevanaGPT
Team project · CapstoneMakes Indian welfare schemes findable in the language people actually speak. Keyword search and 1024-dimension semantic vectors run in parallel, then the two rankings fuse — so a query lands even when nobody knows the scheme's official name. A state-machine chatbot keeps the LLM's behaviour predictable, an eligibility scorer narrows 4,600+ schemes to the ones you'd actually get, and a self-hosted IndicTrans2 service — LoRA-tuned on scheme vocabulary — carries all of it into 22 languages. Four services, one Compose file.
What I builtThe entire backend pipeline — hybrid retrieval (keyword search and 1024-dimension pgvector similarity, fused with RRF), the ingestion and embedding pipeline, the self-hosted IndicTrans2 translation service and its cache, and the four-service Docker Compose orchestration over async FastAPI.
InsightAI
Team project · PrototypeUpload a spreadsheet, get an executive summary, KPIs and charts back — no configuration step in between. An LLM finds the header row first, which is what makes it survive the messy real-world files that break most parsers. The rows go into a FAISS store so you can then ask follow-up questions of your own data in plain English. JWT auth, per-user isolation, and an admin role that can see across accounts.
What I builtThe backend — the FastAPI service, the background file-processing pipeline, retrieval-based Q&A over the FAISS vector store, and JWT auth with per-user file isolation and automated cleanup.
Sign Language Interpreter
Team projectIndian Sign Language to speech, fast enough to hold a conversation on a phone. Classifying hand landmarks instead of raw pixels keeps inference light and steady across changes in background and lighting — a MobileViT front end for the shape of the hand, an LSTM for how it moves through time.
What I builtTrained the classification models — 98.8% on 35 static signs with MobileViT, 95.2% on 6 dynamic signs with an LSTM over 30-frame sequences — and built the Flask server that returns predictions as JSON with confidence scores to the Flutter client, at sub-second latency.
AquaDose
Side build · AndroidA smaller one, and the only thing here that isn't Python. A Jetpack Compose utility that works out how much bleaching powder a well needs, so whoever is doing the chlorination isn't guessing — live calculation, metric or imperial, round or rectangular, against an animated cross-section of the well.
What I builtThe UI and UX — the Jetpack Compose interface, the live-updating result with no calculate step, unit and shape switching, and the animated well cross-section.
I finish things.
Three stages, and the third is the one that usually gets skipped.
Scope it honestly
I'll say when something is a two-hour fix, and when it's past what I know. Both are worth more than an optimistic estimate.
Build it end to end
Data pipeline, model, API, deployment. A system you can open and use, not a notebook that needs someone else to finish it.
Leave it maintainable
Documentation in plain terms, migrations that run, a Compose file that comes up clean. If the next person can't pick it up, it isn't done.
Tell me what you're building.
I'm looking for the next thing to work on. Send the details and I'll reply with an honest read on whether I'm the right person for it.