Cascade
GenAI study companion & hackathon winner
Google GenAI Exchange '24 winner · +30% learning outcomes
GeminiFine-tuningGCPPyQt
AI · Context Engineering · LLM Evals
Harsh Dayal
AI Engineer / Delhi, India
AI Engineer who ships production-grade systems, not demos. I build multi-agent systems, RAG pipelines, and the evals that keep them honest, shipped to production rather than stuck in slideware.
- Agents
Multi-agent systems that ship
- Context
Engineering the right signal
- Evals
Measuring what actually matters
About
I just like making things that work, and making them look like they were built on purpose.
When I was a kid, I spent more time breaking my bicycle down with wrenches than actually riding it. I just wanted to see how the pieces fit together.
That curiosity is how I ended up in software. I didn't have a grand roadmap. I actually got into AI by joining my college's AI club as the graphic designer. But once I was in the room, I wanted to write the logic, not just draw it. I ended up enrolling in a dummy school just so I could skip the classes I didn't care about and spend my time coding.
Most of what I build starts because something in my daily life is annoying, or because someone I know is struggling with a chore. In my second semester, we had a course on Applied Numerical Methods. Instead of memorizing the formulas from a slide, I wrote a Python visualizer to replicate the math so I could study for the final. When I realized I was logging movies on IMDb while my friends were over on Letterboxd, I wrote a sync engine so we could share what we were watching.
For me, if a tool is clunky or painful to use, it isn't finished. I want the things I make to feel like second nature to use, and I want the code behind them to read like prose that someone actually cared about. I also hate guessing. When I was building RAG pipelines and saw everyone arguing about the best chunking strategies, I built BetterRAG to run the benchmarks and look at the real data.
The context is fresh. Bring me something to take apart.
Highlights03
- Winner
- Google GenAI Exchange Hackathon 2024
- Education
- B.Tech CSE, Vellore Institute of Technology · 8.98 CGPA
- Open Source
- Hacktoberfest '24 & '25 · pdfly, MontePy, aegra, rclip, fpdf2
Projects
Systems I designed and shipped, from multi-agent AI platforms to evaluation tooling. Open any one for the brief.
20 / selected
Experience
AI Engineer · Spectatr AINow
Feb 2026 – PresentArchitecting a LangGraph NL2MQL multi-agent system with a 5-layer safety framework and full LLM evaluation + observability.
- Engineered dynamic prompt pipelines injecting complex schemas alongside a 5-layer safety framework for secure, fallback-routed query execution.
- Built a LangSmith evaluation + observability framework across 40+ experiments with 5+ splits: 93.5% accuracy on golden queries, plus user friction/emotion monitoring and regression gates.
- Led an AI dating simulation end-to-end (research to production): parallel LangGraph persona + judge agents on a SCORER framework, reaching 12–20 messages/day retention via a reactive UI.
LangGraphMongoDBLangSmithNext.jsFastAPIAI Engineer (L1) · YoLearn.ai
May 2025 – Jan 2026Owned the core backend for 3 AI Companions and the LLM Ops pipeline behind them.
- Architected backend for 3 AI Companions on a LangGraph LLM–SLM hybrid; implemented Redis-based RAG with persona caching, cutting response latency 82% (P99 <3.5s, TTFT 2.6s).
- Established an LLM Ops & observability pipeline on LangSmith: regression testing across 35+ datasets to track model drift, yielding a 33% performance boost for hybrid models.
- Engineered a Tool Recommendation Engine selecting from 80+ tools via hybrid retrieval, with a CSV to Postgres to Redis pipeline for rapid persona deployment.
- Conducted an internal engineering workshop, 'The Self-Healing AI Dev Loop', converting repetitive manual QA into reusable agent-browser MCP skills, cutting validation cycles from 25 min to under 3 min.
LangGraphRedisQdrantLangSmithPostgresAI Developer Intern · Debales.AI
Feb 2025 – May 2025Deployed 10+ automation agents across support, finance, and sales workflows.
- Multi-agent ticketing system: 40% operational efficiency lift, 15% sales uplift, and 32% engagement increase across support, finance, and sales workflows.
- Engineered NLP chat-tagging at 92% accuracy in the production support pipeline.
- Optimized an Azure Cosmos DB embedding pipeline (HNSW, SHA-256 dedup, quartile chunking) for ~40% storage savings.
LangGraphFastAPIAzureMongoDB
Tech Stack
- Languages
- PythonTypeScriptJavaScriptC++CSQL
- AI & ML
- LangGraphLangChainLangSmithPyTorchTensorFlowScikit-LearnNLPPrompt EngineeringFine-tuning (LoRA/PEFT)OpenAI SDK
- Backend & APIs
- FastAPIDjangoNext.jsWebSocketsMCP
- Databases & Caching
- MongoDBPostgreSQLRedisQdrantDBChromaDBAzure Cosmos DB
- Cloud & Infrastructure
- DockerGoogle CloudAzureAWSVercelRailwayGitHub ActionsCI/CD
- Developer Tools
- GitCursorStreamlitHuggingFaceJupyter Notebook
Certifications
Formal grounding: the courses, specializations, and awards behind the work.
10 / credentials- GitHubJune, 2025
GitHub Foundation
Verify - IBMAugust, 2024
Machine Learning with Python
Verify - NPTELMay, 2024
Cloud Computing
Verify - StanfordApril, 2024
Algorithm Specialization
Verify - Indian Space Research Organisation (ISRO)Feb, 2024
Geodata Processing with Python
Verify - HarvardDecember, 2023
CS50's Introduction to Computer Science
Verify - DeepLearning.AIJune, 2023
Machine Learning Specialization
Verify - MetaJune, 2023
Principles of UI/UX Design
Verify - IBMMay, 2023
IBM Cybersecurity Analyst
Verify - MathWorksMarch, 2023
Machine Learning Onramp
Verify
Writing
A blind benchmark against four deep-research products - DAR won. Here's the architecture.
Jun 30, 20265 min readDeep Research Tools Don't Research. They Summarize. So I Built One Better
Ask four deep research products the same hard question and you get four flavors of the same disappointment. Here's how I built DAR, an agentic research harness that won a blind benchmark against Perplexity, Gemini, Tavily, and Exa.
AgentsLangGraphResearchOpen SourceThe jump from useful to real requires three things almost nobody does.
May 12, 20264 min readContext Injection and Blind Spots: Two Techniques That Make AI Personas Actually Stick
2,000 users in 24 hours. That wasn't the surprising part. Here's how I designed AI personas that feel real, with situation design, trust-gated context, and the blind spot mechanics no one talks about.
AI PersonasLangGraphLLMContext Engineering2+ hours of daily browsing compressed into 15-minute focused reading sessions.
Sep 6, 20252 min readHow I Turbocharge My Day and Knowledge with Perplexity Tasks
How I automated my daily learning routine with 4 simple Perplexity Tasks that deliver curated news, AI research, and system design insights, cutting 2+ hours of browsing into 15-minute focused reading sessions.
ProductivityAI WorkflowLearning
Contact
Let's work together
AI engineer focused on LLM systems, multi-agent architectures, and rigorous evaluation. Always open to ambitious problems, collaborations, and a good technical conversation.
- Based in
- Delhi, India
- Timezone
- IST · UTC+5:30