Skip to content
View rsumit123's full-sized avatar

Block or report rsumit123

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
rsumit123/README.md

Hi, I'm Sumit 👋

Senior backend & data engineer, 7+ years building production systems. I currently lead engineering on an AI-powered brand analytics platform. Outside work I build and ship full-stack AI products end to end: design, backend, infra and deploys. Most of them are live at skdev.one.

Python · FastAPI · PostgreSQL/SQLite · React/TypeScript · Docker · GCP/AWS · LLM agents & tool use · PyTorch


Featured work

AlphaAgent: autonomous LLM trading agents that learn from their own trades · live A fleet of paper-trading agents on NSE (Zerodha Kite) and crypto. Each one reasons with tools, trades within hard risk limits and reflects on every closed trade, and the distilled rules feed back into its next decision. Provider-agnostic LLM engine (Claude, OpenRouter), Langfuse/OpenTelemetry tracing of cost per decision, Google OAuth with Fernet-encrypted per-user keys, and a GCP Spot VM with a watchdog. CI on every push.

Mirth: real-time multiplayer party games for web and Android · live Charades, scribble, trivia and Texas Hold'em over Socket.IO. Live room state lives in memory and is snapshotted to SQLite, so games survive restarts. Includes Quick Play matchmaking with human-like bots, LLM poker seats at three skill tiers, and a 9k-question trivia bank built with a newspaper OCR pipeline. Ships as a Capacitor Android app.

private-llm: a 110M-parameter LLM trained from scratch for about $2 · demo A Llama-style transformer (RoPE, SwiGLU, RMSNorm) pretrained on 2B FineWeb-Edu tokens, then fine-tuned on a custom Indian GK dataset. It beats GPT-2 medium (3× larger) by 19 points on a held-out benchmark.

RunFlow: an AI running coach wired into Garmin · live Syncs runs from Garmin, models fitness and recovery, and pushes structured workouts back to the watch. It exposes everything as an MCP server, so an LLM can review runs and adjust the training plan.

Parakh: food health scores for packaged Indian food · live Scan a barcode or photograph the label. The pipeline goes cache → OpenFoodFacts → vision-LLM label extraction. The LLM only extracts label data and a deterministic Nutri-Score-based function does the scoring, so every verdict is reproducible and explainable.

Chakravyuh: a defence-procurement grand strategy game · live 40 quarterly turns of procurement, R&D and force structure against simulated adversaries, with LLM-written after-action reports. 600+ commits and 250+ tests, with a written decision log for the design choices.

Also

Agora (multi-agent AI debate arena with a read-along voice mode) · ChillBill (Splitwise-style expense sharing, web + Android) · ipl_live (live cricket scorecard API)


🌐 skdev.one · 👤 profile.skdev.one

Pinned Loading

  1. ipl_live ipl_live Public

    live IPL Flask API to get ball by ball scorecard (made by scraping cricbuzz)

    HTML 5 6

  2. chakravyuh chakravyuh Public

    Chakravyuh - Military Company management game (FastAPI + React)

    TypeScript

  3. trade-agent trade-agent Public

    Python

  4. playroom-mirth playroom-mirth Public

    Mirth: real-time multiplayer party games (charades, scribble, trivia, poker) for web and Android

    Python 1

  5. runflow runflow Public

    Self-hosted running coach: imports Garmin and Strava runs, models fitness, generates training plans, pushes workouts to the watch, and tracks adherence. FastAPI + React, with a Claude MCP connector.

    Python

  6. private-llm private-llm Public

    A chat model trained from scratch with nanochat on one H100: training run, benchmarks, Hugging Face demo and serving

    Python