- Replaced LLM-judged scholarship eligibility with a deterministic rules engine at 0.999 precision on 1,382 awards
- Found and fixed a canonicalization bug hiding ~25% of eligible scholarship matches, with zero regressions
- Cut the scholarship reranker's LLM step from 17s to 3.8s median by index-keying its scoring path
- Added rerank and embedding caches so repeat match requests make zero LLM calls: 6.3s to 0.15s in cutover testing
- Halved LLM extraction cost across 40,000+ scholarships by batching pages under one shared prompt
- Isolated failures across parallel LangGraph advisors so one crashing agent no longer ended a student's chat turn
- Stopped deploys silently killing background jobs by moving them onto a durable Postgres queue with a reconciler
- Fixed a production chatbot outage where every streaming reply failed, and removed brittle Gemini fallbacks
Harsh N. Chandak
Full-Stack & AI Systems Engineer
Backends, payments and multi-tenant platforms, and the LLM systems on top
Harsh has shipped two kinds of system. ERP modules that 70+ manufacturers run their plants on, at 20K+ API requests a day, with payments, tax and zero-downtime migrations. And LLM systems where being wrong is expensive, where the real work is an evaluation harness before the feature, a cache, and the canonicalization bug that had been quietly hiding a quarter of the matches for months.
Agent orchestration, running live. The Knight takes the question and decides who can answer it. The Vault Bearer knows him personally; the Flamebearer keeps the record. They read from three stores: the Nameless Temple is home and holds nothing, the Forgotten Shore is the 51 verified bullets he searches, and the Ivory City holds the structured profile.
Ask something. Every answer lists the record it came from.
how it ran
Ask a question. The Knight will go and find it.
Runs entirely in your browser. No API key and no model call, so the answers are assembled only from bullets that exist in a verified bank. It can be wrong about relevance but it cannot make anything up.
Experience
Projects (12)
Job Market Crawler & LLM Ranker
An agentic job-application pipeline that runs itself. 2026
- Built a crawler over 3,500 company career boards on 9 ATS platforms, skipping 86% of 1M+ polls as unchanged
- Surfaced a quarter of new Greenhouse and Ashby postings within 17 minutes of going live, measured on 850 roles
- Collapsed 1,046 duplicate listings of the same requisition and ranked postings deterministically before any Claude call
- Cut LLM cost per document 74% through prompt caching and disabling reasoning on a selection step
- Blocked hallucinated output by rejecting any generated line not found verbatim in a verified source
Multi-Agent AI Pipeline with Evaluation Infrastructure
Production system built at Arizona State University. Aug 2025 - May 2026
- Processed 10K+ audio inputs through diarization, LLM refinement and multi-stage quality evaluation
- Built a 200-input scoring harness, then halved candidates per input while malformed outputs fell ~20%
- Added retry and recovery for recurring model failures, so one bad generation no longer lost a whole run
Developer Telemetry & Observability System
Internal developer tooling built at Arizona State University. Aug 2025 - Dec 2025
- Handled 2K+ events a day with bearer auth, validation and idempotent state tracking
- Cut P95 latency ~45% with async batching and backpressure
- Owned it from event schema through storage model to the React dashboard
Job Alerts + Application Tracker
The first version, born from job-hunt frustration. Summer 2025
- Scraped dynamic job boards on a schedule with Puppeteer and pushed matches to Discord
- Filtered by remote, stack and title before alerting, so the feed stayed readable
- Added JWT auth and multi-user application tracking
Kafka to Neo4j Streaming Pipeline
Live graph analytics under sustained event load. Spring 2025
- Built a Kafka streaming pipeline sustaining 5K events/min at sub-second latency in load tests
- Designed backpressure control to keep ingestion stable under high throughput
- Modeled high-cardinality event relationships as a Neo4j graph, deployed on Docker and Kubernetes
Spatial Data Analysis with Spark and Scala
Hot-spot analysis over large geospatial datasets. Spring 2025
- Ran spatial queries and hot-spot analysis across large geospatial datasets
- Implemented distance and range joins as Spark SQL user-defined functions
Mapping Accident Trends & Patterns
Interactive D3 maps over public accident data. Fall 2024
- Built linked D3 views over GeoJSON so a selection in one chart filtered the map
- Shaped raw public accident data into something a non-analyst could read
Warehouse Robot Optimization using Clingo
Answer-set programming for multi-robot pathing. Spring 2025
- Encoded multi-robot movement, shelf handling and order fulfilment as answer-set programs
- Solved for collision-free plans over a shared warehouse grid
Blockchain & AI for Detecting Financial Data Breaches
Anomaly scoring paired with an immutable audit trail. Fall 2024
- Researched a dual-layer breach-detection architecture using anomaly scoring plus blockchain audit trails
Property Registration System Using Blockchain
A DApp for property transfer and ownership checks. Fall 2022
- Built a property-registration DApp for secure transfers and ownership verification in Solidity and Flutter
OMR Sheets Evaluation Using Image Processing
Grading answer sheets from a phone camera. Spring 2022
- Automated OMR grading with OpenCV and preprocessing tuned for noisy, phone-captured sheets
Real-Time Sign Language Detection
Hand-keypoint classification, live from a webcam. Fall 2021
- Trained a feed-forward network for real-time sign classification on MediaPipe hand keypoints
Education
Arizona State University (ASU)
M.S. Computer Science GPA 4.00 / 4.00
Pune University
B.Tech, Computer Engineering CGPA 8.52 / 10