- 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.
open to SDE / AI roles
Nashik · Pune, India18 months, full-stack SDEERP for 70+ manufacturers
Phoenix, ArizonaMS CS, ASU · 4.0LLM, AI & SDE
drag to spin ⟲
0.999precision on scholarship eligibilityacross 1,382 awards
70+manufacturers run their plants on his ERP modules20K+ API requests a day
~25%of eligible matches were hiddenfound the canonicalization bug
6s → 2sERP dashboard loadjoins and per-tenant caching
17s → 3.8sreranker latency, medianby index-keying the scoring path
zerodowntime on live ERP schema migrationsreversible stages, plants still running
5payment gateways integratedidempotent retries, tax errors to zero
3,547career boards crawled85% of the 28K daily polls skipped
428matches an exact join was dropping419 confirmed by ticker, found by shadowjoin
ask-harsh · 64 verified bullets · answers are instant
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 64 verified bullets he searches, and the Ivory City holds the structured profile.
the Nameless Temple
the Forgotten Shore
the Ivory City
the Vault Bearer
the Flamebearer
the Knight waits at the Nameless Temple
Evidencenothing yet
Ask something. Every answer lists the record it came from.
how it ran
idle
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 (15)
orchestraguard
Agent orchestration never drifts. 2026
- Stopped the failure that costs a parallel run: a backend lead decides user_id, a frontend lead decides userId, both pass their own tests, and nobody finds out until integration
- Refused the second decision at the moment an agent tried to disagree, naming the decision already in force, who made it and who can override
- Decided agreement by comparison and never by a model, with a test that fails if any file in the decision path can reach outside the process
- Escalated worker to lead to PM to human, two attempts a layer, so a deadlock surfaces as a question rather than as silence
TypeScriptMCPNode.js
shadowjoin
Measure the matches your exact join is silently dropping. 2026
- Found 428 matches an exact join was dropping between the S&P 500 and NYSE/Nasdaq listings, 419 confirmed by ticker
- Showed over-normalization costs matches: the full rule set recovered zero on UN country names where two rules recovered five
- Shipped it three ways, as an npm CLI, an MCP server and a Claude skill, so the same scan runs in a terminal or inside an agent
- Ranked the normalization rules by marginal recovery, so the report says which to adopt and in what order
TypeScriptNode.jsMCPPostgreSQL
Job Market Crawler & LLM Ranker
An agentic job-application pipeline that runs itself. 2026
- Crawled 3,547 career boards across 11 ATS platforms, skipping 85% of the 28K daily polls as unchanged
- Caught a quarter of the new Greenhouse postings inside 16 minutes of going live, measured over 806 roles
- Collapsed 1,046 duplicate listings of the same requisition, ranking everything deterministically before any Claude call
- Cut the per-document LLM cost by making a 2,500 token prompt prefix cacheable and killing reasoning on the selection step
- Stopped the model fabricating resume lines by making it return bullet ids, so nothing outside the bank reaches an application
Node.jsMongoDBPlaywrightClaudeLangGraph
source ↗It applies to real jobs with my own resume. The public repo is the engine; the database of applications stays private.
Multi-Agent AI Pipeline with Evaluation Infrastructure
A generative pipeline with the scoring harness built first. Aug 2025 - May 2026
- Treated evaluation as part of the pipeline rather than a step after it: the 200-input scoring harness existed before any tuning, because in a generative system an improvement and a regression look identical from the output alone
- Staged it as diarization, then LLM refinement, then multi-stage quality scoring, so a weak result could be traced to the stage that produced it instead of to the pipeline as a whole
- Made each stage idempotent and lock-guarded, because a run across thousands of audio inputs will be interrupted, and a retry that double-writes is worse than a crash
- Contained model failure at the generation that caused it, so one bad call cost a retry instead of the whole run
PythonFastAPIWhisperXOpenAI APIGemini APIDockerAWS S3
An audio-to-image research pilot at ASU, run over 10K+ inputs.
Developer Telemetry & Observability System
Developer activity events, counted once even when clients retry. Aug 2025 - Dec 2025
- Designed the event schema first and let the storage model follow from it, because a telemetry system that needs a migration to answer a new question stops getting asked
- Authenticated and validated every write and tracked state idempotently, so a client retrying after a timeout could not inflate the numbers it was reporting
- Absorbed bursts with async batching and backpressure rather than a bigger box, so a spike queued instead of timing out
- Carried it through to the React dashboard, so the developers generating the events were the ones who could read them
TypeScriptExpressReactVitePostgreSQLAWS
Internal developer tooling at ASU, ingesting 2K+ events a day.
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
D3.jsJavaScriptNode.jsGeoJSON
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
ClingoAnswer Set ProgrammingPython
Coursework (8)
Graded class projects — a report or a demo where one exists, no deployment behind them.
Arizona State University
- Kafka to Neo4j Streaming PipelineSpring 2025
Live graph analytics under sustained event load.screenshot ↗
- Spatial Data Analysis with Spark and ScalaSpring 2025
Hot-spot analysis over large geospatial datasets.screenshot ↗
- ZK-ColdChain: Private Compliance ProofsFall 2025
Prove a refrigerated shipment broke its temperature limit without showing the log.source ↗
- TrustMed: Drug Knowledge AssistantFall 2025
Answer drug questions by walking a medical graph, not by matching text.source ↗
- Blockchain & AI for Detecting Financial Data BreachesFall 2024
Anomaly scoring paired with an immutable audit trail.
Pune University
- Property Registration System Using BlockchainFall 2022
A DApp for property transfer and ownership checks.report ↗
- OMR Sheets Evaluation Using Image ProcessingSpring 2022
Grading answer sheets from a phone camera.report ↗
- Real-Time Sign Language DetectionFall 2021
Hand-keypoint classification, live from a webcam.report ↗
Education
Arizona State University (ASU)
M.S. Computer Science GPA 4.00 / 4.00
Pune University
B.Tech, Computer Engineering CGPA 8.52 / 10
Tools
Languages
AI / LLM
AI Engineering
Frameworks
Cloud & DevOps
Data