a quick scan for recruiters & humans.
a broadcast tour of how i build.
ai/ml engineer · evidence over vibes
ai · mlops · grounded rag · llm evals
ai/ml engineer bhopal, india
AI/ML engineer who ships with receipts. Model registries, grounded RAG, and the evals that keep LLM systems honest — finished to the point where they're boring to operate. I measure everything I ship.
platforms
ML infrastructure that runs itself
ModelDock — registry, gated promotion, serving, drift watch.
grounded rag
answers that cite their source
Dasaiko — hybrid retrieval, page-level citations, live in prod.
evals
measuring what actually matters
TraceBack — deterministic ground truth vs confident nonsense.
about
I'm a 4th-year CS undergrad at VIT Bhopal. Most of what I build starts the same way: someone says an AI system works, and I ask against what ground truth? That question turned into TraceBack. It turned Dasaiko from a chatbot into a retrieval system with a benchmark suite. It made ModelDock refuse to promote a model that didn't earn it.
I read the papers, then write them up on LinkedIn so they stick.
Off the clock it's two disciplines that keep each other honest. Boxing — the sweet science, where you learn in real time what happens when you overcommit on vibes. And music — I produce and release as aawhan, lo-fi confessionals for the late hours; the Joji-to-Yung-Lean school of feelings. Dilruba, ionknow//idk, Aloof — out on Spotify and everywhere else.
releases — as aawhan · on spotify & everywhere
experience
mar 2026 — may 2026
The company's first web presence, taken from zero to signed-off production.
sep 2026 — oct 2026
Structured ML/AI training track — offer letter on file.
selected work // click a row for receipts
register → version → gate → serve: promotion is earned at the eval-threshold gate, not assumed. 72-test suite in CI/CD, authenticated inference, rate limiting, drift monitoring. MIT.
fastapi · postgresql · redis · next.js · docker
Hybrid retrieval (vector + BM25 + RRF + BGE reranking): +86% recall@5, +75% MRR, +84% nDCG@10 over the RRF baseline. Answers cite the exact page of the source PDF.
react · fastapi · postgresql/pgvector · groq
Grounded baseline: 100% pass rate and root-cause accuracy. Qwen 2.5 3B: 10% on both — at 90% average confidence. 30 runs × 3 scenarios. The gap between fluent and correct is the whole point.
python · fastapi · mcp · ollama · next.js
Investigates production incidents from operational tools, recommends remediation, then verifies recovery before calling it fixed.
python · mcp
YOLOv8 + OpenCV with a dual-layer quality audit — blur detection and confidence filtering — for human-in-the-loop labeling.
python · yolov8 · opencv
LLaMA 3 on Groq, LangChain, Streamlit. No internet search — if it's not in the document, it says so.
python · langchain · groq · streamlit
Detects the track playing on YouTube and opens its Genius page. Small, useful, shipped.
javascript · chrome-extension
Fixed policy-evaluation error logging and strengthened regression coverage. ✓ 498 passed · 33 skipped — verified before merge.
python · context-governance for agentic ai
XML response docs for HealthController endpoints in a .NET 9 aerospace engine simulation platform. ✓ .NET build + Swagger/OpenAPI verified.
c# · .net 9 · swagger
stack
languages
backend & apis
data & ai
databases
cloud & infra
on linkedin // things i'm hyped about
AI/ML engineer focused on grounded systems and rigorous evaluation. Always open to ambitious problems, a good technical conversation, or a beat to collab on.