- Selected through campus placement for a two-month internship.
Bengaluru, IndiaAI / AUTOMATION / PRODUCTS
Vignesh, in motion. ↗Vignesh B S
I make AI useful.
Less busywork. More possibility.
AI systems, automation and the products around them.
Turn the page ↓
Work, over
time.
AI systems, product development
and the testing behind them.
- Build and ship AI features for an EdTech platform: ML pipelines, agent workflows and RAG.
- Own model serving, CI/CD and monitoring in production.
- Lead development of a real-time online assessment platform on Supabase, from research and design through testing and release.
- Tested a live LLM interview platform for accuracy and consistency; wrote automated test plans and tracked latency.
02 / RECOGNITIONBuilt together.
Explore achievements ↗
Built together.
Recognized.
Hackathons, competition results and the learning behind them.
01
Fidelity Investments India Hackathon 2026Winner
02
Hack-Aventus 3.0Winner
03
Infinity Code HackathonWinner
04
Google Maps Platform AwardsCommunity Favourite
02 / PROJECT ROOMStep into
Step into
my projects.
Scroll to move through the work.
VIGNESH B S / AI + AUTOMATIONKEEP SCROLLING ↓
02 / PROJECTSProjects built.
Projects built.
Problems solved.
A field notebook of things I have built. Open a project to see how it works.
01 / 03
SYSTEMS OBSERVATORY / VIGNESH B SKEEP SCROLLING ↓Outsurance
Automating insurance plan comparison.
Only my main projects are listed here. For more projects, see my GitHub ↗
Built to
be useful.
From models to interfaces.
A closer look at what I build,
and what happens under the hood.
Outsurance
Applied machine learningAI Health Insurance Advisor
- A local LLM (Gemma 3) reads unstructured medical history and scores severity, replacing hardcoded rules.
- That score became the top feature (24% importance) in an XGBoost risk classifier (4 tiers, 85% F1, SHAP explanations). A Cosine-KNN ranker then returns the top 5 plans.
Medical history→Local LLM→Risk model→Plan ranking
Gemma 3XGBoostSHAPCosine-KNN
GARUDA
Edge visionEdge AI for Road Safety
- Fine-tuned YOLOv12s on RDD2022 (5 road-damage classes) and built a Damage Severity Score. Regression over past scores forecasts when a road will fail.
- Runs on the edge over existing CCTV and dashcam feeds and sends about 2 KB of JSON instead of video, cutting bandwidth by 99%. Also added YOLO11 with ByteTrack, EasyOCR plate reading and MediaPipe drowsiness detection.
Camera feed→Edge inference→Severity score→Compact JSON
YOLOByteTrackEasyOCRMediaPipe
Sparsh Mukthi
Multimodal AIVoice-Driven 3D AI Classroom
- Students can interrupt the AI teacher mid-lesson. Routes between Gemini, Claude and GPT-4o with strict JSON output.
- Serves Orpheus-TTS (3B) voices through FastAPI and runs MediaPipe in the browser for hand-raise and attention tracking.
Student voice→Model routing→Structured lesson→Natural speech
GeminiFastAPIMediaPipeOrpheus-TTS