Pradumn Patidar — AI Engineer & Data Scientist

About Me

pradumn.md

~/pradumn/about.sh

# Who I am

I'm a Data Scientist & AI Engineer at Tredence Analytics, building production AI systems that bridge generative AI with robust backend engineering — from agentic RAG pipelines to hot-pluggable MCP agents and distributed URL shorteners hitting 1,497 RPS.

# Education

degree: B.Tech. in Metallurgy Engineering and Material Science

institute: Indian Institute of Technology Indore

duration: July 2019 — May 2023

cpi: 7.95 / 10.00

$ cat current_role.txt

Data Scientist @ Tredence Analytics · Bengaluru, India · June 2023 → Present

CAPA Recommendation

Match Accuracy

Req/s Throughput

Async Pytest Tests

MCP Tools Exposed

Work

Experience

Building AI-powered systems in production, at scale.

Data Scientist

Tredence Analytics Pvt Ltd. · Bengaluru, India

June 2023 — Present
Full-time

▸ Autonomous Deviation Intelligence System — LangGraph, AWS Bedrock, OpenSearch, DynamoDB, Terraform

Built a LangGraph pipeline with hybrid retrieval (α=0.45 semantic + β=0.35 BM25 + γ=0.20 metadata), Claude 3.5 Sonnet reranking via Bedrock, and DynamoDB-backed HITL checkpoints; reduced CAPA recommendation from 48 hrs → 23 sec across 13.5K+ deviations. Provisioned all AWS infrastructure via Terraform with modular state management.
Implemented confidence gating with auto-approval thresholds and feedback learning loop using OpenSearch aggregations; achieved 94% match accuracy with 22% precision lift over 500+ validation cycles.

▸ Enterprise Text-to-SQL Platform — LangChain, FastAPI, PostgreSQL, BigQuery, Snowflake, Milvus

Designed multi-database Text-to-SQL supporting 4 databases, 5+ LLMs; 3-level semantic caching (exact → embedding → NER-template), reducing latency by 70% and API costs by 60% across 500+ daily queries. Deployed as a FastAPI microservice with connection pooling and async query execution.
Built SQL Guard validation pipeline (syntax, aggregate, column, join validators) with Jaeger tracing and NER-based disambiguation for multi-turn context retention; integrated automated rollback on validation failure.

▸ Additional Contributions

Built GraphRAG pipeline with Neo4j for HCP network analysis using spaCy NER, co-citation graphs, and Louvain clustering for KOL identification across 50K+ physician records.
Developed anomaly detection using Isolation Forest + Autoencoders on manufacturing signals; reduced batch failures by 12%.

Full Stack Developer Intern

MiBi Services Pvt Ltd. · Remote

May 2022 — July 2022
Internship

Engineered aspirant portal with integrated chatbot for pre-assessment, achieving 500+ user onboarding in two months using Node.js, Express.js, MongoDB.

Featured Work

Projects

Production-grade AI and distributed systems, built from the ground up.

GenAI / Agents

MCP Hot-Swap Multi-Agent Assistant

Hot-pluggable AI agent platform with runtime tool swapping via MCP stdio transport. Static AST scanning catalogs dormant servers without process spawning — new capabilities attach, execute, and detach within a single query lifecycle, zero restarts.

14Tools

4MCP Servers

3Hot-plug Cycles

40+Pytest Tests

$0API Cost

PythonMCP SDKFastAPIRedisDockerOllamallama3.2PrometheusGrafana

Systems Engineering

Distributed URL Shortener

High-performance distributed URL shortening service sharded across 3 PostgreSQL nodes via consistent hashing. Multi-layer Redis caching with Look-Aside + Cache-Null patterns, Redpanda event analytics pipeline, and full Prometheus–Grafana observability.

1,497 RPSThroughput

2.7msP50 Latency

sub-230msP99 Latency

100%Cache Hit

16Services

FastAPIPostgreSQLRedisRedpandaDockerPrometheusGrafanaaiohttpuvloop

Tech Stack