PRODUCTION ARCHITECTURE & PROOFS

Engineered Systems & Proofs.

Deep dives into production distributed topologies: from 19-VM cross-cloud CDC pipelines streaming 24 solar manufacturing tables, to RocksDB stateful deduplication scaling through 21x Flipkart traffic spikes, and adversarial multi-hop red-teaming across 13 mock tools.

SHOWING 5 OF 5 CASE STUDIES
AI & Red-Teaming// 06/2023 – Present
Outlier.AI (Freelance Client)

Frontier AI Evaluation & Multi-Tool Red-Teaming

Adversarial Injection & Multi-Hop Reasoning Stress Testing for LLM Agents

Constructed complex adversarial scenarios across 13 mock enterprise tool environments. Authored deceptive SQL datasets and conflicting state records to probe multi-step agentic reasoning and uncover frontier model failure modes.

13
Mock Tools
Enterprise APIs simulated
4 to 6
Reasoning Hops
Cross-system verification
3
Modalities
Text, Vision & Audio RLHF
40+
Failure Modes
Vulnerabilities discovered
[LLM Evaluation][Red-Teaming][RLHF][Google Gemini 3.5][Mistral 658B][Python][Synthetic SQL][Agentic Scenarios]
ID: claude-code-frontier-ai
Renewable Energy & IoT// 01/2025 – 09/2026
Tata Power Mumbai (via DataSturdy Consulting)

Near Real-Time CDC Platform for Solar Manufacturing

19-VM High-Availability Topology Streaming 24 MSSQL Tables to Databricks

Architected an end-to-end cross-cloud CDC platform streaming 24 MSSQL tables across every stage of solar panel manufacturing. Provisioned a 19-VM HA cluster (Debezium, Kafka, Flink, ClickHouse) in 3 days, recognized by Tata Power leadership.

19
VM Topology
Nodes across HA stack
24
MSSQL Tables
Manufacturing stages streamed
3 Days
Turnaround Time
End-to-end deployment
< 2s
Data Freshness
From factory sensor to OLAP
[Debezium][Apache Kafka][Apache Flink][ClickHouse][MSSQL Server][Databricks][Grafana][Cross-Cloud VPN]
ID: solar-manufacturing-realtime-cdc
FinTech & Trading// 2025
Bajaj Broking (via DataSturdy Consulting)

Voice to Trade: AI-Driven Execution & SEBI Compliance

Automated Audio Extraction & Trade Reconciliation Engine

Architected Voice2Trade for Bajaj Broking: converting call recordings into structured financial orders using Gemini 3.5 Flash and Mistral 658B, with automated SEBI regulatory consent reconciliation.

>98%
Accuracy
Entity & intent extraction
Gemini + Mistral
Model Stack
Flash 3.5 & 658B LLMs
100%
Compliance
SEBI regulatory audit coverage
< 5s
Processing Speed
Per 3-minute trade call
[Google Gemini 3.5][Mistral 658B][Python][Audio Transcription][SEBI Compliance][MSSQL][FastAPI]
ID: voice-to-trade-ai-engine
E-Commerce & Scale// 2025
Flipkart Ads Project (via DataSturdy Consulting)

Flipkart Ads: Near Real-Time High-Throughput ETL

Apache Beam on GKE with RocksDB Deduplication Handling 21x Traffic Spikes

Engineered a low-latency Apache Beam (Java) streaming pipeline on GKE processing millions of ad clicks and views from Google Pub/Sub. Used RocksDB stateful deduplication and Aerospike sub-millisecond cache lookups to handle 21x Big Billion Days traffic.

21x
Spike Tolerance
Big Billion Days peak load
99.9%
Uptime SLA
Automated GKE orchestration
< 1 ms
Lookup Latency
Aerospike metadata enrichment
3
Marketplaces
Flipkart, Hyperlocal, Grocery
[Apache Beam (Java)][RocksDB][Aerospike][Google Pub/Sub][GKE][Apache Spark][Dataproc][ClickHouse][Druid][Azkaban]
ID: flipkart-ads-scale-etl
Enterprise Database// 2025
Infosys (via DataSturdy Consulting)

Infosys ALCON & OMS Database Performance Optimization

Stored Procedure Restructuring & Query Optimization slashes Runtime by 80%

Targeted high-latency legacy Stored Procedures in mission-critical enterprise systems (ALCON and OMS). Implemented strategic indexing and query plan restructuring to achieve a 25% average latency reduction, cutting peak queries from 15 min to 3 min.

80%
Max Speedup
15 min reduced to 3 min
25%
Average Gain
Latency reduction
6
Procedures Deployed
Zero-defect enterprise rollout
[MSSQL Server][T-SQL][Execution Plans][SQL Profiler][Performance Tuning]
ID: infosys-enterprise-db-tuning
All architecture blueprints, benchmark figures, and SLAs are verified against production deployments.
Verify on GitHub (@PraveenBen)