Experience

Where I have worked and what I built there. Newest first.

• 
  2024 to present Founding member and lead engineer

  Perspectivity

  A platform that puts many outlets' coverage of one story side by side, marks where each one leans, and lets a reader question it. I own the backend and the infrastructure. The link resolver, the pipeline orchestrator and the Kubernetes deployment layer all started with me, and I run the production cluster. It is a single $98 VPS.

  The pipeline takes in thousands of articles a day and has ingested more than 400,000. The cluster serves about 58,000 HTTP requests a day to roughly 3,500 distinct client IPs, at a 0.10 per cent 5xx rate, on a three-node MongoDB replica set.

  Every analysed event comes out as an 11-section structured object: source credibility, claim-level fact-checking with three-tier evidence grading, framing bias, context omissions. Where the evidence is missing the schema forces an "Unverified" verdict instead of a guess. There is also an 11-bucket political taxonomy and a 300-constituency election map.

  Above that sits a human-in-the-loop content pipeline with rate caps, an audit trail and auto-publishing to several platforms, on nine production cron schedules.

  It reaches 106,000 YouTube subscribers and 787,000 lifetime video views, and 186,000 web users over thirteen months, peaking at 55,000 in a single month, with readers in 26 countries. Live on the web, with iOS and Android in beta.

  app.perspectivity.codrishtikon.life

• 
  Aug 2025 to Jan 2026 Research and development engineer

  Fellou AI (ASI X Inc.)

  Sole author of three systems for an agentic browser. The first is a generative-UI proof of concept: the model emits a validated UI specification that renders into a live, interactive interface instead of a chat reply. The second runs the desktop browser on a provisioned cloud machine and streams it into a web page, with session lifecycle and authentication fallback.

  Across them, 13 backend services, 4 route groups and 23 custom React components, with typed contracts across the boundary. Work under confidentiality; described here at the level that allows.

• 
  2025 to 2026 Technical lead

  AIMarketCap

  Backend lead on a seven-person team, leading four people and reporting to the CEO. The work covered the data platform, how far it scales, the analytics on top of it, and the interface and product design.

  About 15 external data sources feed it: crypto, equities, AI-tool directories, startup lists, social signals, web traffic. The time-series engine computes rolling metrics across 11 timeframes, from 15 minutes to yearly, for roughly 20,000 coins, behind a market-bubble interface.

  The AI-tool corpus runs to about 75,000 records. Collecting it took a six-month campaign against bot detection, with undetected browser drivers, CDP mode and rotating proxies.

  I also built a schema-embedding retrieval system. It reads the database schema, embeds each collection into a Milvus vector index, searches that index per question, writes an aggregation pipeline and phrases the answer. Three embedding providers are pluggable, with Milvus and pgvector behind it.

  All of it runs on a self-managed kubeadm cluster with a private registry: 238 Kubernetes manifests and 18 Dockerfiles, including 23 network policies, 18 stateful sets, 12 pod disruption budgets and 9 Prometheus service monitors, with Airflow orchestrating the crypto, stock, AI-tool and web-traffic pipelines.

• 
  2026 to present Technical lead

  Flannel.ai

  A consumer product for household records: deeds, warranties, vehicle and health documents. Built under a SOC 2 framework.

• 
  2024 to 2025 Engineer

  Freedoms.ai

  Built the whole backend for an AI notes and journaling product. Retrieval runs over the user's own notes with LightRAG and Language Agent Tree Search. Speech is self-hosted in both directions, tuned to the platform rather than called from a managed API. I developed and evaluated several open-source speech models to choose between them, at a time when there was very little evaluation work for speech.

• 
  Feb to Jul 2024 ML research and open-source intern

  Menlo Park Lab

  Sole engineer on an autonomous coding agent built with LangChain and LangGraph. It planned its own steps, wrote and tested code, and ran shell commands, closing the loop without a person driving each step. The founder adopted it internally.

• 
  2024 to present Research lead

  BanglaLLM

  Open research on language models for Bangla. Three published papers, a safety benchmark, 224,000 released instruction examples and an open model family from 1B to 13B parameters, plus the data tooling behind them.