Scroll
01
Sec – 01

About

Creating efficient AI systems and building them at scale. I enjoy personalising models and making tools adaptive and more productive.

Focus

The model isn't the most interesting thing. It's the systems around like custom LoRA models, data pipelines, feedback loops, custom harnesses, and deployment.

02
Sec – 02

Work

  • Better Campus, SWE & Applied AISoftware EngineerBuilding agentic pipelines, products and gamification for 2 million users across all platforms. Worked with Browserbase for better QA.Better Campus
    • Full Stack
    • LangChain
    • Agentic Pipelines
    • Postgres
    • Cloudflare
    • Browserbase
    Repository↗
  • In Depth Context Workflow for AgentsSoftware EngineerFalse Claims Act — processed 100GB of document data with custom context graph, retrieval and embedding pipelines for case-building agents.Operis
    • Context Graphs
    • LangChain
    • LLM
    • Embeddings
    • Retrieval
    • Benchmarking
  • Manga Net (Stable Diffusion)A zero-LoRA over animagine-xl-4.0, trained on 15k custom images captioned with Qwen2.5-VL, to generate more natural manga panels. Context graphs and novel breakdown keep panels consistent, with a dual IP-Adapter holding character identity. DM if you're curious about this work.LoRA
    • LoRA Models
    • Eval
    • Fine Tuning
    • Stable Diffusion
  • PerceptCalHacks 12.5 WinnerNeuro-guided video optimisation with TRIBE v2, predicting and visualising population-average cortical response from a brain activation matrix.TRIBE v2 model
    • Model Compression
    • Cloud Deployment
    • Redis
    Repository↗
  • Flagship-LLM-Royale (10,000 views)A low-latency pipeline where a YOLO model reads the board and hands it to an LLM as a coordinate matrix. The model makes its own decisions and plays against other flagship models.YOLO, Computer Vision
    • YOLO
    • Fine Tuning
    • Vision
03
Sec – 03

In Progress

  • Stable DiffusionManga Net 2The second version of Manga Net wider caption better tuning and tighter eval loop aimed at panels that hold their character across a page rather than one frame at a time. These models need understanding other than characters...
  • CompetitiveGrinding Codeforces to CMover half way there...
  • AgentCoachAn agent that teaches math on a whiteboard, drawing as it works through a problem. Anthing dropped on the canvas becomes context the agent reasons overs. Gift for my brother :D
04
Sec – 04

Archive

  • BlueprintTurns a hackathon URL into ten original ideas, then checks an idea you already have against what has been built before.Semantic search
    • Python
    • Embeddings
    • Retrieval
  • Insider Correlation ToolkitInsider trades correlated against forward returns at +7, +14 and +30 days, with a backtesting harness attached.Quantitative
    • Quantitative
    • Backtesting
    • Pandas
  • routeqAn evaluation harness for code and client quality — whether a change improved anything or just moved the failures out of sight.Evaluation
    • TypeScript
    • Evals
    • Quality
← All portfoliosRobotics →