8.9 → 5.0 s
Median time to first token with prompt caching
Median time to first token per turn before and after transcript caching on a ~80k-token loop (4 tool rounds) Per-turn range: 4.6–22.0 s before 3.4–5.1 s after
Vinayak
AI Engineer
I build agents that run real tools and cite their sources
Neo is guiding youTap anywhere to stop
About
I’m an AI Engineer at AMnova Technologies architecting full-stack multi-agent systems with FastAPI LangGraph and Next.js
My focus is practical reliability building agents with sandboxed code execution source-grounded citations and deterministic guardrails that prevent hallucinations
Featured projects
All systems & archive (14)- Sandboxed execution The Code agent gets a real Linux box (shell filesystem web search dev servers) and streams every tool call over a typed FastAPI WebSocket contract into a live trace and preview
- Grounded retrieval Learn turns PDFs web pages and YouTube transcripts into answers with [n] citations and graded courses and abstains when the sources don't cover the question
- Resilient routing Each iteration is classified and routed across eight models
8.9 → 5.0 sMedian time to first token with prompt caching
- LangGraph
- FastAPI
- Next.js
- E2B Sandbox
- pgvector
- Ingest and deduplicate Aggregates 9 job sources validates apply links and deduplicates postings on canonical URL plus company title and location
- Transparent ranking Ranks jobs with local MiniLM embeddings and SQLite FTS5 across freshness source quality location role fit and skills so every match score can be inspected
- Guardrails and security Rejects tailored CV output that fabricates numbers or requirements
9Job sources merged and deduplicated
- Next.js 16
- React 19
- TypeScript
- SQLite + FTS5
- Drizzle ORM
- Explainable ML + Voice
ARIA
Predicts which customers will leave then explains why in plain words and voice
- Imbalanced classification XGBoost trained on 10,000 customer records with SMOTE for class imbalance and one scaler shared by training and inference
- Risk tiers and drivers FastAPI serves real-time churn probabilities grouped into Low Medium and High risk alongside global feature-importance rankings
- Explanation and voice Llama 3.3 70B writes plain-language explanations and next-step retention strategies; ElevenLabs speaks them
10kCustomer records trained on
- XGBoost
- SMOTE
- scikit-learn
- FastAPI
- Llama 3.3 70B (Groq)
- Three disease models An SVM for diabetes a Random Forest for stroke and Logistic Regression for heart disease each in its own scikit-learn pipeline
- One FastAPI service A single FastAPI backend serves the web interface the prediction endpoints and the AI assistant with no frontend build step containerized with Docker
- Contextual explanations Patient inputs and model outputs go to Llama 3.1 8B on Groq which explains the result in plain language and answers follow-up questions
3Clinical screening models
- scikit-learn
- FastAPI
- Llama 3.1 (Groq)
- Docker
- Hugging Face Spaces
- Eight-agent pipeline Ingest clean EDA visualize ML benchmark insights code gen and chat run in turn on one uploaded CSV streamed live over WebSockets
- Local ML fallback pandas scikit-learn and XGBoost compute every numerical result locally so provider outages never break the core analysis
- Model benchmarks Four ML models are benchmarked on every run and the core analysis needs no LLM calls
8Autonomous pipeline agents
- FastAPI
- WebSockets
- React
- pandas
- scikit-learn
- XGBoost
- Gemini
- Groq
Loom.ai
One AI workspace to chat code in a live sandbox study and delegate
- LangGraph
- FastAPI
- Next.js
- E2B Sandbox
- pgvector
- Groq
- OpenRouter
- WebSockets
github.com/vinayak533/loom.ai
8
Models routed with automatic provider fallback
Models in the routing table across Groq OpenRouter and OpenCode as listed in the README A failed call retries on at most two alternates
10
Specialist agents with human approval
Specialist agents on one parameterised graph each pausing for human approval as listed in the README
PROBLEM
Chat windows can't execute code or check answers against private documents Coding research and study each need their own tools and a real environment to run in
MECHANISM
LangGraph agent in an E2B sandbox; turns routed across 8 models retried on up to 2 alternate providers
SOURCED CLAIMS · CLICK ONE TO HEAR IT
Chatbots talk about code Loom runs it inside a real Linux sandbox
Every turn is routed across 8 models If a provider fails the turn falls back to the next one
Caching the transcript cut the median time to first token measured turn by turn
And the limits are written down too
route_turn() · try it: take a provider down
turn
classified
classified
Each provider toggles up/down Primary first then up to two alternates
Provider names from the project; the replay is illustrative
time_to_first_token() · median ~80k-token test loop
No cache8.9 s
Cached transcript5.0 s
Measured values from the performance method doc; playback is 10× faster than real time
CLICK A STAGE:
Turn Classifier & Router
Routes across 8 models (Groq OpenRouter OpenCode) with 2-hop provider fallback
LangGraph Checkpointer
Persists graph state after every node visit; caches ~80k-token transcript
E2B Sandbox & pgvector
Concurrent read-only tool batches; sequential writes; lexical hashed RAG
INPUTS
User Prompt / Task · Repo Filesystem · PDF · Web · YouTube Sources
OUTPUTS
Typed WebSocket Trace · Live Dev Server Preview · Cited [n] Notebook Answers
Sandboxed execution
The Code agent gets a real Linux box (shell filesystem web search dev servers) and streams every tool call over a typed FastAPI WebSocket contract into a live trace and preview Read-only tool batches run concurrently; writes run in order
Grounded retrieval
Learn turns PDFs web pages and YouTube transcripts into answers with [n] citations and graded courses and abstains when the sources don't cover the question Retrieval uses lexical hashed vectors in pgvector
Resilient routing
Each iteration is classified and routed across eight models On a rate limit quota or 5xx it retries on up to two alternate providers and LangGraph checkpoints let sessions survive cold starts
STACK (8 TOOLS)
Orchestration
Backend & Data
Model / API Layer
Runtime & UI
DOCUMENTED LIMITS (3 OF 26)
- Rate limiting is in-process
- The checkpointer defaults to SQLite on local disk
- One concurrent run per session
LazyHire
Nine Kerala job boards one ranked list no duplicates
Here’s a simulated run of the CV guard Pick the edit you think it rejects
tailor_cv() · pick an AI-suggested edit
source CV (fictional)Built a FastAPI service that parses supplier invoices into PostgreSQL
One of these edits invents something The guard will catch it
Toy guard: it checks numbers only LazyHire's real checks cover numbers and requirements
github.com/vinayak533/lazyhire
LazyHire pulls 9 job sources around Kerala's tech parks into one ranked list minus the duplicates
Every match score opens up: freshness source location role fit and skills
It tailors your CV and rejects any edit it can't trace back to the original
The code is public: read it on GitHub
- Job sources merged and deduplicated
- 9
- Job sources the ingest step aggregates deduplicated on canonical URL plus company title and location
- Application stages tracked
- 4
- Stages in the app's application tracker
- Private guarded CV review
- PDF · DOCX
- CV formats the private review accepts
ARIA
Predicts which customers will leave then explains why in plain words and voice
A toy dial not the trained model: drag it to see how the risk tier changes
score_churn() · illustrative toy not the trained model
Low risk
p(churn) ≈ 12%
Low risk Activity is recent so the usual check-ins are enough
github.com/vinayak533/bank-churn-intelligence-platform
ARIA predicts which bank customers might leave: XGBoost on 10,000 records balanced with SMOTE
It sorts them into Low Medium and High risk and surfaces the top five drivers
Then Llama 3.3 explains why in plain words and ElevenLabs reads it aloud
- Customer records trained on
- 10k
- Rows in the bank-customer training dataset per the README
- Actionable risk tiers
- 3
- Churn-probability bands from the README: Low ≤35% Medium 35–65% High >65%
- Top predictive features surfaced
- 5
- Global feature importances from the trained model shown in the UI They rank the model's drivers overall not per customer
PredictLab
Screening results that explain themselves
Preprint · Zenodo
PredictLab published with a DOI
DOI 10.5281/zenodo.20253956
Zenodo record v1.0 17 May 2026 A preprint not peer-reviewed
github.com/vinayak533/PredictLab
PredictLab screens for diabetes stroke and heart-disease risk from 30 health inputs
Three models an SVM a Random Forest and a Logistic Regression behind one FastAPI service
Llama 3.1 explains every result in plain language Vinayak published it as a preprint
- Clinical screening models
- 3
- An SVM (diabetes) a Random Forest (stroke) and a Logistic Regression (heart disease) each a scikit-learn pipeline behind one FastAPI service
- Health inputs analysed
- 30
- Input fields across the three screening forms: 8 + 10 + 12
Skills & technologies
What I build with
Each tool links to the project or role that used it.
Python for the AI and ML work, TypeScript for the Next.js front ends, and SQL for the data underneath.
- Python
- TypeScript
- SQL
- C++
- Java
Agents that call real tools, routed across models with fallback, retrieval that cites its sources, and guardrails that reject made-up output.
See it in use in Loomscikit-learn and XGBoost for tabular models, feature engineering on skewed data, MLflow to track experiments, and pgvector for retrieval.
See it in use in AmdoxFastAPI services over REST and WebSockets, Next.js on the front, and agents that run code inside E2B Linux sandboxes.
See it in use in LoomPostgreSQL and Supabase for data, SQLite where it runs locally, and Docker and CI/CD to ship to AWS.
See it in use in AMnova
Professional experience
Where I’ve worked
- Jun 2026 – PresentLLM store-audit pipeline
AMnova Technologies
AI/ML Engineer · Kochi, IndiaROLE & RESPONSIBILITIESAI/ML engineer on the product team. I work on the LLM pipeline that checks new merchant stores and the backend behind the merchant dashboard.
KEY CONTRIBUTIONS & IMPACT · SELECT ONE FOR DETAILSBuilt the LLM pipeline that checks new merchant stores at AMnova: It speeds up the store audits that were done by hand during merchant onboarding.
- Python
- LangChain
- FastAPI
Wrote the FastAPI + Supabase backend behind the merchant dashboard: Its queries run async.
- FastAPI
- Supabase
Packaged the inference services in Docker: GitHub Actions deploys them to EC2 on every push.
- Docker
- AWS EC2
- GitHub Actions
- Apr – May 2026Forecasting and churn models
Amdox Technologies
Data Science Intern · Bengaluru, IndiaROLE & RESPONSIBILITIESBuilt retail demand-forecasting and customer-churn models and served them as FastAPI endpoints.
KEY CONTRIBUTIONS & IMPACT · SELECT ONE FOR DETAILSBuilt demand-forecasting and churn models: XGBoost and scikit-learn on transaction histories.
- Python
- XGBoost
- scikit-learn
Served the models behind FastAPI endpoints: Other apps call them over REST.
- FastAPI
Tracked experiments in MLflow and built Streamlit dashboards: So the business team could see the results.
- MLflow
- Streamlit
- Aug – Sep 2025Credit-approval classifiers
Cognifyz Technologies
Data Science Intern · RemoteROLE & RESPONSIBILITIESBuilt credit-approval classifiers on a heavily imbalanced financial dataset.
KEY CONTRIBUTIONS & IMPACT · SELECT ONE FOR DETAILSTrained and tuned credit-approval classifiers: Checked on F1, precision, recall and ROC-AUC, not accuracy alone, because the classes are skewed.
- Python
- scikit-learn
Feature engineering, outlier cleaning and class-imbalance handling: On skewed applicant data.
- pandas
- Feature engineering
Certifications & achievements
The Record
Zenodo, DOI 10.5281/zenodo.20253956.
Participant, HackerRank Orchestrate Hackathon
Built and submitted a terminal-based autonomous support-triage agent.
Top 2%, LeetCode Weekly Contest 487
Rank 108 of 6,770.
Bachelor of Computer Applications (BCA)
Mahatma Gandhi University, Kerala
Advanced Data Science & AI Specialization
Avodha Centre for IT Education
7-month intensive program, accredited by NSDC and Skill India.
- Fundamentals of Agents
- Claude Code 101
- AI Fluency: AI Capabilities & Limitations
- AI Fluency: Framework & Foundations
- Agents and Workflows
- Introduction to Generative AI: Art of the Possible
- Complete Generative AI Course: RAG, AI Agents & Deployment
- SQL (Advanced)
- Python Coder
- Data Visualisation: Empowering Business with Effective Insights
- Deep Learning using TensorFlow
- Machine Learning with Python
- Data Analysis with Python
- SQL and Relational Databases 101
- The Data Science Course: Complete Data Science Bootcamp 2025
NeoOnline
Try asking
Voice conversationHands-free · talk over Neo any time
Tap me to talk
