SVS PraveenSeshanagottu Venkata Sujith Praveen
I build AI systems that route intelligently instead of guessing. Turning production-grade RAG, multi-agent orchestration, and scalable backends into systems that actually ship.
Type any query or keyword to route directly to Praveen's portfolio data.
An AI engineer who thinks in systems, not demos.
I keep building the same pattern at different scales — a router that reads intent and dispatches to the right specialist instead of forcing one model to do everything. You'll see it in a multi-agent system that routes financial queries to specialized analysis agents, and in the very intent router at the top of this page. It's not a coincidence; it's how I think about hard problems — break them into the smallest correct piece, then route to it deliberately.
I care as much about what happens when things go wrong as when they go right. Every system I ship fails closed, not open — from role-based access and jailbreak-resistant guardrails in a clinical RAG system, to a hallucination-detection layer that scores every AI-generated answer before it reaches a user. My focus is architecting cross-domain AI applications and scalable backend systems that hold up under real conditions, not just demo conditions.
A stack built for production-grade AI applications.
Agentic AI & Orchestration
Retrieval & Search
LLMs & Vector Databases
ML & Evaluation
Full Stack, Backend & Automation
Cloud, Deployment & Security
Hands-on work across AI pipelines and scalable Python systems.
AI/ML Engineer Intern
MobcoderAI, Noida, UP (Onsite)- Improved SEC-filing retrieval accuracy from 62% to ~95% (NDCG@5) by building a hybrid BM25 + dense-vector search pipeline with Cross-Encoder re-ranking on Qdrant Cloud.
- Reduced query response time to under 5 seconds across 100+ SEC 10-K/10-Q filings by designing a multi-agent orchestration system in LangGraph with parallel routing on Llama 3.3 70B, using MCP for standardized tool-calling between agents.
- Lowered LLM evaluation costs significantly while maintaining high faithfulness scores by building a deterministic RapidFuzz + PyTorch scoring pipeline tracked in MLflow.
Agentic AI, governed RAG, and production systems.
SPrav Job AI (Pro v2.4)
An autonomous, local-first AI career engine that scans 28,700+ tech company ATS portals (Greenhouse, Lever, Ashby, Workday, SmartRecruiters) and Wellfound, calculates cosine similarity against master PDF resumes, and prepares STAR-tailored applications for 1-click dispatch.
- Scans and monitors 28,700+ real tech ATS company career portals and job feeds with <18ms SQLite query latency and <120MB memory footprint.
- Engineers dual-stage matching: dense Sentence-Transformers embedding cosine similarity + multi-criteria LLM rubrics to score genuine role fit.
- Generates tailored STAR cover letters and outreach pitches with a 1-click human review dispatch queue, supporting local Ollama models and cloud inference (Gemini / Groq).
- Shipped as a 100% private, air-gapped standalone Windows portable application (v2.4 Pro) with zero-trust local XOR credential encryption.
RespiRAG
A production-grade, zero-hallucination RAG system for Non-Small Cell Lung Cancer (NSCLC), built for real oncologists, pharmacists, and clinical nurses.
- Built a governed RAG pipeline over 600+ FDA drug labels and 3,900+ pages of NCCN clinical guidelines, combining HyDE query expansion, MedCPT biomedical embeddings, and Cross-Encoder re-ranking to reduce hallucinations.
- Implemented role-sensitive access (Patient/HCP/Admin) with JWT-based RBAC, refresh-token rotation, Redis session blacklisting, and a regex-based guardrail layer against prompt-injection attempts.
- Reduced redundant vector-DB and LLM calls with a Redis citation cache and multi-key Groq rotation with Ollama fallback for graceful rate limiting.
- Built a CI/CD-compatible offline evaluation harness and a document version-control system that flags superseded clinical literature.
Finance RAG Copilot
An enterprise-grade, multi-agent RAG pipeline that analyzes SEC financial filings (10-K/10-Q) with zero hallucination.
- Built a multi-agent router (Financial Analyst, Comparison, Decomposition, Risk Analyzer, Trend Agent) that classifies each query and dispatches it to a specialized agent.
- Combined dense and BM25 sparse retrieval with Reciprocal Rank Fusion and cross-encoder reranking against Qdrant Cloud for accurate multi-document financial search.
- Engineered a zero-cost deterministic self-evaluation layer scoring Faithfulness, Groundedness, Hallucination Rate, and Citation Accuracy on every response — no LLM judge required.
- Designed a 5-state visual guardrail system that flags unverified numbers, warns on low-confidence answers, and hard-blocks out-of-scope questions.
Governed-RAGFlow
A production-grade AI visual orchestration engine combining a drag-and-drop workflow canvas with an asynchronous Python backend, featuring real-time prompt-injection & PII guardrails, semantic multi-agent routing, and self-healing retry loops.
- Built a full-featured visual drag-and-drop pipeline builder in React 18 with live step-by-step execution telemetry, topological graph sorting, and interactive pre-built onboarding templates.
- Engineered real-time fail-closed governance guardrails (PII masking, toxicity filtering, and regex/LLM prompt-injection defenses) that short-circuit malicious inputs before LLM execution.
- Implemented an adaptive self-healing retrieval architecture with semantic fallback routing that dynamically switches knowledge branches upon detecting low-confidence or irrelevant documents.
- Integrated PostgreSQL (Neon) persistence for comprehensive audit trails, token telemetry, latency logging, and deterministic faithfulness evaluations.
ShipForesight
An AI-powered logistics platform that predicts shipment delays before dispatch using a 3-stage ML pipeline, with plain-language multilingual explanations.
- Built a 3-stage zero-inflated pipeline: an ensemble classifier (CatBoost + LightGBM + XGBoost) predicts delay risk, a regressor estimates delay days, and a reason classifier explains why.
- Achieved 84.7% accuracy / 0.891 AUC-ROC on delay prediction, with a DuckDB feature store enriching predictions using real vendor on-time-rate and route history.
- Integrated live multi-modal routing (Truck, Ocean, Air, Train) with real routing APIs, including automatic detection of non-viable routes like train-across-ocean.
- Generated plain-language, multilingual (English, Hindi, Marathi, Gujarati, Tamil) explanations of each prediction via LangChain + Groq.
These are the ones I'm proudest of — there's more experimentation happening in my repos.
Explore More on GitHubReal client, real payments, delivered solo.
Ajeevan Samarpan Parivar — NGO Website & Donor Platform
Independently designed and built a full-stack website for an NGO from the ground up — starting with a 5-day throwaway prototype to meet an urgent launch deadline, then rebuilding it as a production React application with a Node.js/Express backend over the following weeks.
- Integrated Razorpay for donation payments, handling real transactions end-to-end.
- Used MongoDB for structured contact-form and donation logging.
- Built a custom admin dashboard where the NGO could view and manage every contact submission and donation record without needing developer help.
- Delivered solo, end-to-end, in under 3 months — from prototype to production.
Continuous learning across GenAI, Cloud Architecture, and Security.
Cloud Certifications
Web Development & Full-Stack Certifications
Security & Identity Certifications
Cisco Ethical Hacker
Cisco Networking Academy
Certificate document will be linked soon.
IBM Cybersecurity Analyst
IBM
Certificate document will be linked soon.
GitHub activity and contribution graph.
Let's build intelligent systems that recruiters remember.
Open to AI/ML engineering roles, Python backend opportunities, GenAI internships, RAG systems work, and applied machine learning collaborations.
Quick Chat / Discussion on AI Applications
Fill out the form below or reach out directly at svspraveens@gmail.com.