About
Production war stories, with the numbers attached.
I'm Rahul, an AI engineer with five years of building products end to end: backend platforms, data pipelines, and the LLM systems on top. Everything I ship has to survive real users, real latency budgets, and real bills. Most of what I learned doing that isn't written down anywhere, so this site is me writing it down.
What I've shipped
- An agent platform that loads skills on demand. Tools stay locked until the agent reads the playbook for the job in front of it. Runs in production inside a WhatsApp-native assistant.
- A glucose forecaster built on a Transformer-LSTM hybrid, predicting spikes two hours ahead with 94% of predictions inside Clarke Zone A. The LLM extracts features. It doesn't tell fortunes.
- A vision pipeline for meal logging that handles ten thousand photos a day at a p90 of about six seconds, and lifted in-app logging from 24% to 38%.
- An agent, judge, and extractor eval pipeline that scores production traces with compaction in the picture, because a benchmark that ignores how context degrades is measuring the wrong thing.
- Prompt-cache engineering across multiple providers. A single timestamp in a system prompt once bought us a 0% cache hit rate. Once.
- A Text-to-SQL agent with deterministic scoring. No AI in the math.
Where I've worked
Canvas AI 2022 to present
Founding software engineer. First engineer on an agentic investing suite, a consumer-health assistant serving thousands of users a day on a streaming chat backend, and a WhatsApp business-intelligence agent running NL-to-SQL at better than 90% accuracy. Built the company's analytics engine from scratch: funnels, retention, cohorts, and segmentation on a dynamic query builder over ClickHouse, fed by a real-time event SDK and eight data connectors. I also built the eval and observability tooling the other products lean on, plus the infrastructure underneath: Kafka streaming, Airflow orchestration, Dockerized microservices on AWS, full CI/CD. The unglamorous parts that keep the glamorous parts honest.
Quantile Analytics 2021 to 2022
Quantitative analyst. Automated rebalancing-signal pipelines across twenty-plus global indices, and an NLP parser for SEC filings that replaced a lot of manual spreadsheet archaeology.
Education
B.Tech from IIT Indore, class of 2021.
The fascination started there. Somewhere in the middle of my degree I took Andrew Ng's machine learning course on Coursera and fell hard for it, less for the demos than for the math underneath. Backpropagation, gradient descent, the perceptron. One assignment had me coding a CNN, and I still remember the small shock of watching it learn.
Receipts
- JEE Main 2017: rank 3,289 out of 1.2 million. Top 0.27%. JEE Advanced: rank 4,473.
- CAT 2021: 99.19 percentile. Chose to keep building instead.
- Bronze medal, BitGrit data science contest, Inter-IIT Tech Meet 2019.
What this site is
A place to share my work, what I'm learning, and the occasional war story. When a post makes a claim, there's a repo attached that backs it up.
The name comes from Cyberpunk 2077. The Blackwall is the firewall that keeps rogue AIs away from everyone else's net, which felt about right for someone who ships agents for a living. The domain is a Gibson reference. Agent memory is the problem I care most about, and Neuromancer got there forty years early.