AI-ML Engineer · Multi-Agent Systems · Applied Engineering Architectures
Production agent systems · Grounded document intelligence · Eval-driven reliability
Tumakuru, India · UTC+05:30 · Open to remote / India full-time
| Area | Stack Badges |
|---|---|
| Languages | |
| AI / ML | |
| Backend & Systems | |
| Frontend | |
| Infra & Tooling |
I design and ship resilient, scalable AI/MLOps systems that survive contact across the complex environments.
My work lives at the intersection of:
- Multi-agent orchestration with adversarial stress-testing & recovery paths
- Spatial & grounded RAG for engineering documents
- Enterprise voice AI with failover-aware reliability
- Long-term memory systems with fact reconciliation
- Classical ML pipelines that deliver measurable business impact
I care deeply about observability, failover, deterministic evaluation, and architectural honesty.
🛡️ AEGISAutonomous Executive & Geospatial Intelligence System Hierarchical multi-agent command center
Stack: FastAPI · LangGraph · LangChain · Gemini · PostgreSQL + PostGIS + pgvector · Next.js · Tailwind · Leaflet · Docker |
Enterprise Voice Platform Three modules under one resilient architecture:
Stack: TypeScript · React · Express · FastAPI · Hunar Voice · SQLite · Tailwind |
|
P&ID Revision Intelligence + Grounded Spatial RAG Industrial document intelligence for engineering drawings.
Stack: TypeScript · Python · React · Vite · Express · Gemini · Groq · Docker · Vercel |
🌙 mayaTerminal AI Companion with Long-Term Memory Research-grade CLI companion focused on memory integrity.
Stack: Python · LiteLLM · SQLite · NumPy · Pydantic · pytest |
| Project | Focus | Signal |
|---|---|---|
| Smart Loan Recovery System | Default risk + recovery strategy | Random Forest + K-Means · 85% Acc · 89% ROC-AUC · Streamlit app |
| GoodFoods | Multi-location reservation agent | Tool calling · 75+ restaurants · Groq + Llama 3.3 |
| apartment-price-prediction | Layout-based valuation | End-to-end regression pipeline |
| Cricket-Data-Analytics | Sports analytics | Exploratory + predictive notebooks |
| small-language-model-from-scratch | LLM fundamentals | Training a small LM from first principles |
| discover-the-transformer-architecture | Architecture deep-dive | Transformer internals |
I optimize for resilience over convenience and auditability over black boxes.
| Principle | Practice |
|---|---|
| Fallback everywhere | LLM routing, telephony, search providers |
| Facts have lineage | Supersede — never hard-delete |
| Stress before ship | Adversarial plans before execution |
| Observability first | Tracing, token cost, confidence scores |
| Evals ship with product | Harnesses are not optional |
- Production multi-agent systems, FDE / applied AI roles
- Industrial & engineering AI (document intelligence, spatial systems)
- Real-time voice & telephony AI
- Long-horizon memory & agent reliability work
Open to collaborations, research discussions, and high-impact engineering roles.
| vinayks705@gmail.com | |
| vinay-k-s-4b203024a | |
| GitHub | Vinay-014 |
| Location | Tumakuru, India · UTC+05:30 |
Built with adversarial resilience. Optimized for systems that hold up.