LEVEL 01NEXT IN QUEUE
FLAGSHIP · RAG PLATFORM
IntelliRAG
Answers need receipts. Trace the passage, test the claim, and refuse when the docs fall short.
- PINNED SOURCE EVIDENCE
- 20 SUPPORT SCENARIOS
- 54 RAG CHECKS
- PUBLISHED FAILURES
SYSTRIS · FOLLOW THE THREAD
THE HUMAN INTERFACE A CONNECTED BODY OF WORK
A MAP OF THE WORK
Pick a star for the whole story, or click the central Tetris piece for GitHub. The full galaxy connects every chapter.
EXPLORE THE FULL GALAXY ↗RAG · MEMORY · EVALUATION · INFERENCEGITHUB PROFILE ↗
COGNITIVE ENGINE AUTO PLAY
THINK · PLACE · MAKE ROOM
Curiosity finds the gap.
Ask a better question.
0000000 LINES CLEARED
NOW
I
NEXT IDEA
T
PLAYER 1 > SYSTRIS : TETRIS WITH SYSTEMS@huesofbanter · Charan RathoreCharan Rathore · I build systems that make the pieces click.
Analyst at MiQ, ex-Flipkart - I build RAG, memory and eval systems.
PUBLIC GITHUB ACTIVITY · LIVE
The work keeps moving.
A September-only replay of public merged PRs, ending with the month at a glance.
Checking current public GitHub activity...
SEPTEMBER 2026 · REPLAY THE MERGES
Source: GitHub public merged PRs in September 2026, checked 26 Sept 2026, 12:00 am (last known result; live check unavailable). Blocks follow event order, not a date-spaced scale. Red means a merged PR classified as a fix or test by its title. Pushes and CI results are not included.WHAT BROUGHT YOU HERE TODAY?Pick a question.
Pick a question.
Find the starting layer.
Choose a piece. Follow the decisions down to the work.
HERE, RIGHT NOW · ISTToday
━ Main thread ─ Supporting connection
THE STORY STARTS WITH YOU
What pulls you in?
An answer you can trust. A pattern worth following. Something that feels easier to use. Choose what you came looking for.
Line width highlights the main story, not a performance score. Every connection is curated from the work.
THERE’S MORE WHEN YOU’RE READY.
STAGE 02 · THE PROJECT STACK
Watch the work fall into place.
Start with the systems I am building now. Follow the code, try the live work, then explore the rest.
Python
FastAPI
Celery
PostgreSQL
Redis
ChromaDB
OllamaLEVEL 02NEXT IN QUEUE
MEMORY ENGINE · LIVE
memoRABLE
Six memory blocks. Every useful line stays linked to its source.
TypeScript
Next.js
pdf.jsLEVEL 03NEXT IN QUEUE
FORECASTING · REPOSITORY
Time-Series Temperature Modelling
Forecasts meet rooftop truth. Bias checks and a leaderboard show what holds up.
React
MLflow
LightGBM
IoTLEVEL 04NEXT IN QUEUE
FULL-STACK · LIVE
Project Management Tool
A shared board for the messy middle: drag, decide, move.
TailwindLEVEL 05NEXT IN QUEUE
MULTI-AGENT AI
Agentic Finance Advisor
Agent research with a visible confidence check, not a mystery verdict.
LangChainLEVEL 06NEXT IN QUEUE
COMPUTER VISION
Drone Wildlife Detection
Drone footage into blackbuck detections. Computer vision with a field job.
YOLOv8
PyTorchLEVEL 07NEXT IN QUEUE
CAD AUTOMATION
3D Bolt Dataset Automation
Generate the shape and its label together. 1,000+ bolts ready for CAD ML.
THE HUMAN BEHIND THE SYSTEMSCuriosity in.
Curiosity in.
Something useful out.
Analytics taught me to look for the signal. Building makes me test whether I understood it. Here’s the loop I’m trying to get better at.
MOVE 01 / NOTICE
Find the real question.
- What comes in
- A behavior that doesn’t make sense.
- The choice
- Start with what someone is trying to do. In seller analytics, that means following the journey before deciding what to measure.
- What I keep in mind
- A clean dashboard is useful only if it changes a decision.
SIDE QUEST · FROM THE BLOGThe two seconds
The two seconds
you never see.
You press Enter and an answer appears. I spent months on what happens in between. Stretched out, those two seconds are exactly one Tetris line: ten cells, and the model only turns up for the last five.
what actually happens after we press enter?
0.00sof 2.00s · waiting on you
IDENTITY
Who just pressed Enter?
Before anything can be remembered or retrieved, one question has to be answered. The login screen you scroll past is solving the hardest problem on the internet, and it isn't intelligence. It's identity.
- IDENTITY
- 0–200ms10% of the wait
- MEMORY
- 200–600ms20% of the wait
- CONTEXT
- 600–1000ms20% of the wait
- THE MODEL
- 1000–2000ms50% of the wait
Three invisible systems, then the model. Half the wait is the part everyone credits; the other half is the part that makes it possible.
STAGE 03 · PLAYER STATS
One commit. Another block.
The board keeps growing. Explore the daily stack, language pieces, and the next repositories in the queue.
I really love designing systems that love me back :)
“Most tutorials show you how to call an API. I want to understand how to build the infrastructure that makes AI work reliably at scale.”
I care about correctness over convenience, documented tradeoffs over hand-waving, and systems that fail gracefully when things go wrong.
Evaluating LLMs for Zeolite Synthesis Event Extraction
A systematic look at prompting strategies for scientific event extraction — what general-purpose models can (and cannot) pull from experimental procedures.
READ ON ARXIV ↗Analyst / MiQ
Working across MENA markets.
OPEN RESUME ↗Product Analyst Intern / Flipkart
Seller funnel analytics on the search personalization team. Pipelines and dashboards turning behavior into product decisions across millions of daily searches.
OPEN RESUME ↗Independent AI Systems Builder / Personal Projects
Production-grade systems from first principles: RAG, multi-agent AI, computer vision, quantitative finance. Documented tradeoffs, measured failure modes.
OPEN RESUME ↗1 CREDIT IN THE TRAY
INSERT COIN.
LET’S BUILD SOMETHING.
Product analytics, AI systems, or any technically interesting problem. Usually reply within 24 hours.
