Open to AI/ML & SWE internships

Abhishek Bevinkatti


FINAL-YEAR B.E. CS (AI & ML) · JAIN COLLEGE OF ENGINEERING, BELAGAVI

Building/shipping Systems that are usefull & measurable.

Retrieval pipelines, multi-agent orchestration, and the evaluation harnesses that keep them honest. A few of the systems I've shipped — from a 0.927 RAGAS research assistant to a browser-native PDF editor that never touches a server.

View work Say hello →
0.927 RAGAS SCORE· 38% LATENCY REDUCTION· <15ms P99 RETRIEVAL· 90%+ FRAUD DETECTION ACCURACY· 12 PDF TOOLS, ZERO SERVERS·
About

Somewhere between a CS degree and a habit of building things.

The non-resume version.

I'm a final-year CS (AI/ML) student who somehow ended up running a handful of real, deployed systems off of free-tier infrastructure and pure stubbornness. It started as "something concrete to Ace the Resume" and turned into an actual habit — I'd rather spend a weekend building something as open source project then prepare for the exam on tuesday. which can be done on monday evening.

Most of what I build starts from half cooked ideas of new project ending with taking my weekdays an ends evolving into solid projects. hence I also lean hard on measuring things — RAGAS scores, latency numbers, false-positive rates and quality in them.

Outside of that: way too many side projects ideas lying in notes, and a genuine soft spot for Dogs (ik this was outta nowhere)

Off the clock
  • Based in Belagavi, Karnataka
  • finding a jugaad to make something work within my constraints that breaks some budget constraints
  • Strong, unsolicited opinions about font embedding
  • Will not ship a number I haven't measured myself
  • also Prefers songs in the background while building/coding etc
Selected Work

The Solid recent ones, all held to the same standard —
does it work, and can you prove it ?

Every project ships with the numbers attached.
Nothing here is a demo you can't measure.

Production

Multi-Agent AI Research Assistant

A RAG platform that scores its own answers before it gives them to you.

LangGraph orchestrates parallel research agents across a FAISS retrieval layer, with self-correcting evaluation loops and per-response RAGAS scoring built in — not bolted on. Real-time SSE streaming and structured latency, quality, and error tracking, deployed live on Hugging Face Spaces.

LangGraphGroq · Llama 3.3 70BFAISSFastAPIStreamlitDocker
Live demo → GitHub →
0
RAGAS Score
0
Faithfulness
0% ↓
Latency (6.5s → 4.0s)
<0ms
P99 Retrieval
Open Source

Loop Watchdog

A kill switch for AI agents that won't stop retrying.

A local-first FastAPI proxy sitting in front of OpenAI-compatible coding agents, scoring traffic across five signals — retries, file churn, error signatures, edit-oscillation, request growth — and blocking runaway loops before tokens burn. A Cloudflare Worker control plane pushes HMAC-signed alerts to Slack and email.

PythonFastAPITypeScriptCloudflare WorkersD1pytest
GitHub →
0
Detection Signals
0+
REST Endpoints
Blocks on HTTP 409 before the next retry fires.
Open Source

PDFZero

A 12+ tools PDF editor that never sends your file to a server.

Text editing, merge, split, compress, redact, unlock, OCR and more — 100% client-side. A font-classification engine maps 40+ font families for pixel-accurate text edits, with offline OCR via Tesseract.js and a code-split bundle under 42KB on first load.

ReactVitePDF.jspdf-libTesseract.jsZustand
Live demo → GitHub →
0
PDF Tools
0%
Client-side
0+
Fonts Mapped
<0KB
Initial Payload
Complete

Fake Job Postings Detector

Flags a scam listing before you waste an application on it.

TF-IDF feature extraction and ML classification identify suspicious postings with explainable, confidence-scored risk insights — wrapped in a Streamlit interface built for fast, real-time checks.

Scikit-learnNLPStreamlit
GitHub →
~0%+
Detection Accuracy
0% ↓
False Positives
Complete

StyleSync Elite

Fashion generation tuned to run on a GPU you can actually afford.

Stable Diffusion XL Turbo optimized for low-latency inference on consumer-grade hardware, with a FastAPI backend and a prompt-controlled generation pipeline for real-time fashion outputs.

Stable Diffusion XLPyTorchFastAPI
GitHub →
5–7s
Inference Time
0GB
VRAM Required
In Development

RAG Evaluation CLI

Benchmark any RAG system without hitting an external API.

A framework-agnostic CLI for deterministic RAG evaluation — schema auto-detection for JSON/JSONL/CSV, RAGAS-style metrics, and multi-system comparison for faster iteration.

PythonCLIRAGAS
GitHub →
Next up

Publishing to PyPI, then a public benchmark comparing three open-source RAG stacks head to head.

The System

Not a skills list — the layers things actually run on.

Grouped by what each layer does,
the way the architecture diagrams look.

Orchestration, Agents & Reasoning
LangGraphLangChainAgentic AIMulti-agent workflows
Retrieval & Evaluation
FAISSRAGASSentence-Transformers
Serving & APIs
FastAPIREST APIsAsync / SSE streaming
Models Run
Llama 3.3SDXLGemmaGroqGemini
Ops & Observability
DockerPrometheusGrafana
Core Languages
PythonJavaJavaScriptC / C++
Timeline

Recent internships, then the degree behind them.

Experience
AI & ML Intern
Edunet Foundation × IBM SkillsBuild — Remote
DEC 2025 — JAN 2026
Cybersecurity with GenAI Intern
Vodafone VOIS × Edunet Foundation — Remote
DEC 2025
Education
B.E. Computer Science (AI & ML)
Jain College of Engineering, Belagavi
EXPECTED JUN 2027
Get in Touch

Seeking AI/ML & SWE roles.

I'm looking for AI/ML and SWE internships — happy to talk through the projects, what I can bring to your team, or we can just skip ahead and :
Salary discuss karle ?

Reach-out EMAIL → projectsGITHUB → Connect/say hi LINKEDIN →