// building the future, one agent at a time
VISHAL
VISHAL
PANJETA
SMTS | ML Engineer | AI Architect
|
Scroll
About Me
I build things that think. Over the past 8+ years, I have gone from writing NLU engines at a small startup to architecting multi-agent systems at ThoughtSpot that answer the kind of messy, real-world business questions most pipelines struggle with.
I have shipped patient insight platforms in healthtech, NLU engines in fintech, and AutoML systems in enterprise SaaS. The common thread: taking a hard ML problem, building something that actually works in production, and making it fast enough.
0+
Years in the Trenches
0
Companies Built For
0
Sectors
HealthTech, FinTech, Enterprise SaaS
GitHub Activity
Loading contributions…
Experience
ThoughtSpot
- Designed and built a multi-agent orchestration framework with dynamic expert agent registration. Agents plug in dynamically into the system.
- Built an AutoML pipeline from scratch: model selection, training, evaluation on business data. Fine-tuned transformers handle forecasting and anomaly detection.
- Cut average API latency by 30% with custom caching layers optimized for time series and language workloads.
- Built a "why" engine: grounded data analysis that explains KPIs, powered by a multi-agent pipeline under the hood.
Wellthy Therapeutics
- Owned the patient insight system end-to-end: meal, activity, and health goal analysis for thousands of patients daily under low bandwidth constraints.
- Architected the backend from scratch in TypeScript/Node.js/Express. It handles 60%+ of all API traffic.
- Rebuilt the chat platform on MQTT and XMPP. Throughput jumped 80%.
- Pushed platform SLA past 99.5% using AmazonMQ and AWS SQS.
- Built the platform SDK for secure third-party integration of Wellthy's system into external platforms.
BankBuddy.ai
- First engineer on the team. Built the product from ground up.
- Wrote NLU engines using Rasa and custom lightweight models for fintech conversations.
- Flask API engine with multithreaded pooling, sustained high throughput for investment banking clients.
- Fast inference pipelines for mission-critical banking algorithms where latency = money.
Intel Corporation
- Built Jenkins CI/CD pipelines across multiple EC2 instances. Build times dropped 20%.
- Made a code flow analysis tool that auto-generates documentation and Postman collections. Saved developers a significant amount of documentation time.
Projects
LLM-MRI
- Prompt explorer with tokenization, attention heatmaps, and next-token probabilities.
- Interactive modules for 3D embeddings, layer-by-layer activation views, and hallucination scenarios.
LLM Memory Compression Lab
- Six interactive modules covering memory decay, compression tradeoffs, retrieval accuracy, and architecture comparisons.
- Context window strategies and graph memory growth visualized with D3 and drag-and-drop controls.
PyLLaMa-CPU
- Python bindings on top of llama.cpp for CPU inference.
- 4-bit quantization support, with 7B running at high throughput on CPUs.
- Works with ggml/gguf model files.
What If?
- Generates alternate history timelines from a single counterfactual prompt.
- Streamlit UI with Docker deployment.
- Grounded plausible alternate history generation using LLMs.
ModiScript
- Esoteric, Turing-complete language built from PM Modi's quotes.
- Programs start with "mitrooon", with no floats, imports, or comments for "Make in India" vision.
- Includes a web UI and CLI for running .chai files. Built the lexer, parser, and interpreter with Python ASTs.
Skills & Tools
AI / ML
LLMs
Transformers
RAG
Multi-Agent Systems
AutoML
NLU / NLP
Computer Vision
Deep Learning
Anomaly Detection
Forecasting
Backend
Python
FastAPI
Node.js
TypeScript
Flask
Express
Cloud & Infra
AWS
Docker
Kubernetes
Jenkins
Redis
Kafka
Lambda
API Gateway
Azure
GCP
Terraform
Data
PostgreSQL
Time Series Analysis
MQTT
XMPP
AmazonMQ
SQS
RabbitMQ
