Saivamshi
Jilla
Building intelligent systems, not just models.
Ask Saivamshi Anything
What I've Built
Every project told as: Problem → Approach → Tradeoffs → Result
AgentBay
AI Agent Marketplace · Jan 2026 – Present
Food Package & Freshness Detection
Computer Vision System · Nov 2024 – Dec 2024
Technical Arsenal
Not a list. A map.
How I Think
This is the section most portfolios skip. Engineering philosophy — not just what I built, but how I decide.
RAG vs Fine-tuning — My Framework
I default to RAG when knowledge is external, frequently updated, or needs source traceability. Fine-tuning earns its place only when the task is highly specialized, data is abundant, and latency is critical. Mixing both — RAG with a fine-tuned retriever — is usually the real-world answer.
Evaluation-first Engineering
Before writing a single line of ML code, I define what 'good' looks like. Precision vs recall tradeoffs, latency budgets, hallucination thresholds. At LinCXO, the 85% accuracy target wasn't arbitrary — it was reverse-engineered from what reduced manual effort by 60%. Metrics drive architecture.
On Agents and Trust Boundaries
An agent is only as reliable as its tool contracts. I design agents with explicit failure modes: what happens when a tool times out, returns null, or hallucinates an action. In AgentBay, sandbox isolation isn't just security — it's a trust boundary that makes agents composable.
Shipping is the Proof of Work
A model that scores 95% on a benchmark but hasn't touched production data is a demo. I optimize for deployment from day one — FastAPI-wrapped, Docker-containerized, observable. The pipeline matters as much as the model. That's what distinguishes ML engineering from ML research.
AI Engineer InternCurrent
- ›Independently delivered a multilingual (Hindi/English/Marathi) AI answer-sheet evaluation system — Gemini 3.5 + multi-engine OCR (Mathpix + Google Vision), 85%+ accuracy across 500+ verdicts
- ›Designed a self-learning feedback loop from evaluator corrections, with guardrails and OCR fallback / confidence-based routing
- ›Orchestrated an 80+ node n8n workflow, cutting manual-review rate to 17%
- ›Built idempotent MySQL schemas + Python ingest services for audit-complete, traceable evaluation storage powering client dashboards
- ›Developed AI-integrated features for ADTech products: LLM content generation, intelligent targeting, automated campaign optimization
AI Intern
- ›Built LLM-powered resume + JD parsing system achieving 85%+ accuracy, reducing screening time by 60%
- ›Deployed Playwright web scraping pipelines to aggregate job listings at scale
- ›Developed candidate scoring logic enabling bias-reduced, data-driven hiring decisions
- ›Built LLM agents with function calling to automate internal hiring workflows
- ›Prototyped AI video generation agent from text prompts
AI Intern (Remote)
- ›Processed and prepared 5GB+ dataset using end-to-end data cleaning pipelines
- ›Reduced delivery timelines 25% using IBM Watson cloud-based NLP workflows
- ›Deployed chatbot that increased user engagement with educational AI content
B.E. Computer Engineering
Thought Stream
ML concepts, distilled. 4+ articles on Medium.
10 Essential ML Algorithms Explained: Decision Trees to Neural Networks
Beyond Accuracy: Master ML Evaluation Metrics
Feature Engineering Mastery: Transform Raw Data into ML Gold
ML Lifecycle Explained: Problem to Production (Why 90% Fail)
10 Essential ML Algorithms Explained: Decision Trees to Neural Networks
Beyond Accuracy: Master ML Evaluation Metrics
Feature Engineering Mastery: Transform Raw Data into ML Gold
ML Lifecycle Explained: Problem to Production (Why 90% Fail)
10 Essential ML Algorithms Explained: Decision Trees to Neural Networks
Beyond Accuracy: Master ML Evaluation Metrics
Feature Engineering Mastery: Transform Raw Data into ML Gold
ML Lifecycle Explained: Problem to Production (Why 90% Fail)
Let's build something
remarkable.
Open to AI/ML roles, research collaborations, and interesting problems. Graduating May 2026.
saivamshijilla04@gmail.com