Amit Patil
Final-year ECE student with 6 months of backend & ML experience at Wayfair. Built a production AI incident triage bot using a hybrid RAG pipeline (pgvector + BM25 + GPT-4o). Looking for software & ML/AI roles where I can ship reliable, data-driven systems.
About
I'm a final-year Electronics and Communication Engineering student at BMS Institute of Technology, graduating in 2026 with a CGPA of 9.34.
Over the last 6 months at Wayfair's Global Availability Platform team, I got to work on real production systems — debugging Kafka-based inventory streams, running load tests for critical APIs, and building GAP-Bot: an AI incident triage copilot that I owned end-to-end from schema design to Kubernetes deployment.
I like working at the intersection of backend engineering and applied ML — building systems that are reliable, observable, and actually useful to the people on call at 2am.
When I'm not writing Python or debugging Datadog dashboards, I'm probably reading about distributed systems, tinkering with FPGA boards, or thinking about the next thing to build.
Experience
Wayfair
Software Engineering Intern — Global Availability Platform
Bengaluru, India
GAP-Bot: AI On-Call Triage Copilot
Designed, built, and deployed end-to-end — an AI-powered Slack bot that auto-triages inventory incidents using a hybrid RAG pipeline (pgvector + BM25 + RRF + GPT-4o). Owned schema, retrieval algorithms, agent orchestration, prompts, CI/CD, and GKE deployment independently.
Bharat Electronics Limited
Trainee Intern — PDIC
Bengaluru, India
Projects
GAP-Bot
AI On-Call Triage Copilot
A Slack Socket Mode bot that auto-triages on-call inventory incidents at Wayfair. Detects new Jira/Slack tickets, runs a two-stage hybrid RAG query over 161+ historical incidents, enriches results with live Datadog signals, and posts a structured triage card — root cause, implicated service, investigation steps, and reference tickets — in ~11 seconds.
- ▸Two-stage hybrid retrieval: pgvector cosine ANN + PostgreSQL BM25, fused with Reciprocal Rank Fusion (k=60) and a 4-signal composite reranker (semantic 0.40, BM25 0.30, recency 0.15, source 0.15)
- ▸Multi-step agent orchestrator with per-step tracing, execution memory, and graceful per-step failure isolation
- ▸Datadog triage engine with YAML-driven L0–L4 pattern router and parallel log queries (6 workers)
- ▸Deployed on GKE with Cloud SQL proxy sidecar, Vault secrets, Buildkite CI/CD with CodeQL SAST
- ▸Anti-hallucination guard, multi-provider LLM abstraction (GPT-4o / Gemini / Claude), cost estimation
YOLO on FPGA
ML Model on PYNQ Z2 to Control a Robotic Arm
Deployed a YOLOv3 object detection model onto a PYNQ Z2 FPGA, achieving a 3× speedup in image processing vs. CPU. Used a hyperspectral image database to detect fruit ripeness stages for robotic arm control.
- ▸3× speedup in image processing time vs. CPU baseline
- ▸Hyperspectral image database for fruit ripeness detection
- ▸Hardware-software co-design: FPGA acceleration + Python control layer
Twitter Data ETL Pipeline
Python · Airflow · AWS EC2 & S3
Built an end-to-end ETL pipeline to extract tweets from the Twitter API, transform data using pandas, and load results into AWS S3. Orchestrated workflows with Airflow DAGs on EC2 with automated scheduling, retries, and basic monitoring.
- ▸Automated Airflow DAG scheduling with retries and failure alerting
- ▸pandas-based transformation layer before S3 load
- ▸Deployed on AWS EC2 with S3 as the data lake destination
RepoAtlas
Local-First Developer Intelligence for Unfamiliar Codebases
A locally-hosted static analysis tool that helps engineers onboard onto unfamiliar codebases fast — interactive dependency graphs, API route discovery, and composite risk scoring, all without code ever leaving the machine.
- ▸Interactive dependency graph (React Flow) with circular-dependency and blast-radius detection
- ▸Auto-discovers FastAPI, Flask, and Express routes; 0–100 risk score per file with plain-English explanations
- ▸Fully local-first — FastAPI + SQLite backend, Next.js dashboard, Docker Compose deploy
ContextBench
Does Semantic Context Actually Improve AI Agent Accuracy?
A benchmark that isolates whether semantic context genuinely improves LLM agent accuracy on enterprise data queries — or whether gains are just an artifact of more tokens — using a synthetic 5.5M-row / 42-table warehouse and a token-matched control tier.
- ▸Control tier matches token count of the richest context tier but with generic filler — isolates signal from noise
- ▸40 hand-written business questions with verified ground-truth SQL and plausible wrong answers
- ▸8 intentional data traps (soft deletes, grain collisions, timezone bugs); graded programmatically via SQL diff, not LLM judgment
Visual AI Agent
Privacy-First Browser Activity Intelligence
A Chrome extension (Manifest V3) that builds a searchable, timestamped record of browsing activity — periodic screenshots, OCR + rule-based summarization, all running fully offline by default with an optional OpenAI opt-in.
- ▸Pluggable AI provider architecture — local rule-based + OCR by default, opt-in OpenAI for richer summaries
- ▸Searchable timeline dashboard (React + Vite) with session filtering and screenshot playback
- ▸Privacy-first: no raw keystrokes or form/paste content captured, one-click disable toggle
Research
Evaluation of Modern Deep Learning Architectures for Automated Dental Malocclusion Detection
Amit Patil, Deepa N Reddy — Dept. of ECE, BMS Institute of Technology and Management
Benchmarked ConvNeXt and Swin Transformer against a ResNet-50 baseline for classifying dental malocclusion (healthy/overbite/underbite/crossbite) from 297 labeled radiographs. Swin Transformer reached 93.44% accuracy (vs. ConvNeXt 86.89%, ResNet-50 83.62%), a statistically significant gain confirmed via McNemar's test with Bonferroni correction across 5-fold cross-validation, alongside a morphological feature analysis (arch angle, symmetry index, aspect ratio) and a computational cost comparison across architectures.
Skills
Education & Achievements
BMS Institute of Technology and Management
BE in Electronics & Communication Engineering
Relevant Coursework
Microsoft Certified: Azure AI Engineer Associate
Official Microsoft certification validating skills in building, managing, and deploying AI solutions on Azure.
2nd Place — PBL Open Day
Won 2nd place at BMS Institute of Technology for the project "Ultrasonic Parametric Array".
CTF — Paperless Pursuit
Attended the "Paperless Pursuit" Capture the Flag event hosted by Accorian on 7th May 2024.
Bengaluru, Karnataka
Open to full-time roles in software engineering, backend, and ML/AI — in Bengaluru or remote.