Systems
Incognito Text
Simple, minimalistic, anonymous texting web application.
- Chat rooms and anonymous accounts
- Flask backend with relational database
- Deployed on Heroku and AWS (PostgreSQL, SQLite)
Experience
5 roles
Where I have worked
Markers
- Systems work
- Machine learning work
- Both
- Systems work: Designed and shipped a fault-tolerant repository sync service (VPC-isolated Lambda, EFS, EventBridge, CDK) that mirrors repos of up to 1,800 branches every 5 min and self-corrects failed pushes, ending stale-source builds that each cost a day of release time.
- Systems and machine learning work: Built Python tooling that diffs patched and unpatched test runs across 200-300 MB logs, separates new regressions from flaky failures and reruns only failing tests; with LLM agent automation on top, this halved manual effort per patch cycle.
- Systems work: Own EMR on EC2 security patching end to end, from builds and release manifests to multi-wave rollout across commercial, GovCloud, China and sovereign regions; shipped 467 patches across 26 release lines, cutting patch lag from months to monthly. Also worked on moving EMR release builds (x86, ARM64) to a new, reproducible build system.
- Systems work: Led resolution and root-cause analysis for 20+ production incidents, including a critical kernel CVE fix for a major bank shipped in ~1 day and a customer-facing AWS Glue 5.0 rollback across 3 regions mitigated in ~4.5 hours.
- Systems work: Launched AWS Glue 6.0 in 5 regions and drove its multi-wave GA rollout; moved its container image builds into the CI/CD pipeline, removing a manual release step.
- Systems work: Cleared 200+ operational tickets in 3 weeks (team queues ~250 to 115 open); mentored 2 engineers to run patching on their own.
- Systems work: Designed and built an event-driven backend workflow (DynamoDB Streams, SQS, Lambda) that checks EMR Serverless vCPU limit requests against account risk scores, EC2 capacity and per-cell Fargate utilization, then raises the quota across 30+ regions, cutting turnaround from 31 days to minutes.
- Systems work: Made stream processing idempotent under at-least-once delivery (retries, backoff, dead-letter queue); extended the DynamoDB schema and API endpoints for ops portal views of decision factors and gated manual approvals.
- Systems and machine learning work: Built AWS data pipelines that collect traffic and pedestrian data for Washington, D.C., and trained a graph transformer on ~20M traffic-jam records from 4 years to forecast road-segment traffic (R-squared 0.86).
- Systems and machine learning work: Built an LLM entity extraction system for emergency call transcripts and a web app that shows each call on a timeline.
- Machine learning work: Built a real-time traffic incident detection system using brain-inspired (spiking) neural networks, and an LLM-based (RoBERTa, GPT) ontology-mapping system; the work produced a Springer paper and a PCT patent.
Projects
7 featured, 9 more
Things I have built
Intelligence
Hate Speech Detection Using Multi-LLM Architecture
Classification using an ensemble of fine-tuned LLMs with in-context learning.
- Two-tiered multi-LLM architecture using LLaMA models (1B, 3B)
- 18% faster processing and 54% lower memory usage
- Fine-tuning with Torchtune and ICL
- Proposed: model calibration, LoRA/QLoRA for efficient fine-tuning
Systems and intelligence
Waze-Based Traffic Forecasting Using Graph Transformers
City-scale traffic prediction using STGformer (graph transformer).
- Trained on Washington D.C. Waze data (4 years, ~20M jams)
- Engineered road network graph; sparse subgraph sampling in PyTorch
- Achieved R² ≈ 0.86
Systems
Maze Solver Using AI
Heuristic and brute-force based maze solver.
- Pygame GUI with random maze generation
- Solvers: BFS, DFS, A* (Manhattan/Euclidean)
- Execution time analysis across 30 random mazes
- A* mean 0.01s; BFS 0.029s; DFS 0.06s
Intelligence
Goal-Oriented Image Quality Assessment Using CNN
Novel image quality assessment powered by deep learning.
- Task-specific image quality evaluation
- ResNet-based depth estimation on NYU-Depth-v2
- Pearson correlation 0.89 and ~1900% efficiency improvement
Systems and intelligence
LLM vs LLM: An Approach to Fault Localization
Adversarial LLM framework for software fault localization.
- Gemma-3-12B generates buggy C++; a second model localizes faults
- Custom fault injector for 5,000 C++ programs
- QLoRA fine-tuned debugger improves accuracy from ~20-30% to ~55-65%
- Demonstrates PEFT feasibility
Intelligence
Single-Lead to 12-Lead ECG Conversion Using RNN
LSTM network predicts multiple ECG leads.
- Predicts remaining 11 leads from a single lead
- Noise removal: Butterworth filter and DWT (db-6 wavelet)
- Experiments with noise removal strategies
- Peak R² 0.9 (some leads); peak average R² 0.56
More repositories (9 of 16)
- Image Recognition Using PCA
- Custom Image Filtering
- Before You Sign
- Divide N Conquer
Filecoin track winner at HackNITR 4.0
- BlockBoard
Winning team at HackVerse Tamil Nadu, a state-level Web3 hackathon
Education
2 degrees
Where I studied
University of Virginia
Master of Computer Science
- Teaching Assistant, Discrete Math and Theory
IIITDM Kancheepuram
B.Tech (Honors) in Computer Science and Engineering
- Gold Medal for All-Rounder of the Graduating Batch (509 graduates)
Toolkit
24 tools
What I work with
Languages
- Java
- C/C++
- JavaScript
- SQL
- Bash
- Python
Cloud and systems
- AWSEC2, Lambda, Step Functions, ECS/Fargate, EKS, DynamoDB, SQS, S3, EFS, IAM, VPC, CloudWatch, CDK, CloudFormation
- Amazon EMR
- AWS Glue
- Apache Spark
- Docker
- Linux
- Git
Engineering
- distributed systems
- CI/CD
- infrastructure as code
- monitoring and alarming
- incident response
- unit and integration testing
ML and AI
- PyTorch
- TensorFlow
- scikit-learn
- LLM fine-tuning (LoRA/QLoRA)
- AI coding agents and MCP tooling
Publications
1 paper, 1 patent
Written down
Brain-Inspired Traffic Incident Detection for Effective Communication
M. Saravanan, Sravanth Potluri
Detects traffic accidents in CCTV footage with brain-inspired networks (spiking neural networks and Legendre Memory Unit RNNs), then uses predictive quality of service on V2X data to pick the network nodes that relay the alert to emergency services.
Network Node and Method for Ontology Mapping in a Communication Network
Ontology mapping for telecom networks: maps the concepts of a legacy product ontology onto a new one with semantic similarity, using causal relationships between objects to get the mapping right. Application PCT/IN2024/050495.
Awards
5 awards
Recognition
Gold Medal for All-Rounder of the Graduating Batch, IIITDM Kancheepuram (509 graduates)
Award for securing the 3rd-highest CGPA in the 1st year of B.Tech across all branches
Part of the winning team at HackVerse Tamil Nadu, a state-level Web3 hackathon, for BlockBoard
Repository for Part of the winning team at HackVerse Tamil Nadu, a state-level Web3 hackathon, for BlockBoardFilecoin track winner at HackNITR 4.0, for Divide n Conquer
Repository for Filecoin track winner at HackNITR 4.0, for Divide n ConquerIIITDM Vashisht 2024 Capture the Flag, 2nd place
Contact
Email first
Write to me
A builder who puts intelligence into systems, and enjoys the hard problems that come with it.
sravanth.chowdary.potluri@gmail.com