Skip to content

Skills

Skills

Grouped by what the work actually is, with the thing that demonstrates it. No self-assessed percentages — those are made up, and you can check everything below.

4

Languages shipped in

11

Projects built

10

Upstream PRs merged

//Areas[7]

  • RL & Agents

    Building the environment, not just the policy.
    • RL / RLVR workflows
    • Reward & verifier design
    • Task generation
    • Long-horizon environments
    • Agent evaluation
    • Trajectory analysis
    • Failure-mode analysis
    • Gymnasium
    • Regression coverage for exploits

    Where it shows up — Snorkel AI production environments; Bespoke Labs pod lead; VERITAS.

  • Interpretability

    Causal methods, not correlational stories.
    • Mechanistic interpretability
    • TransformerLens
    • SAELens
    • Sparse autoencoders
    • Activation patching
    • Causal interventions
    • Probing classifiers
    • Ablations
    • Representation engineering
    • Activation steering

    Where it shows up — GPT-2 trajectory paper; AXON; BrainPatch.

  • ML Systems

    Making a model run fast enough and reproducibly enough to trust.
    • PyTorch
    • Hugging Face Transformers
    • vLLM
    • llama.cpp
    • KV cache / prefill / decode
    • Continuous batching
    • LoRA / PEFT
    • Quantization
    • MoE routing
    • Model sharding
    • Mixed precision
    • Gradient accumulation
    • Checkpointing
    • Model packaging

    Where it shows up — BrainPatch runtime across Transformers, llama.cpp and vLLM; Noma at ~16 ms.

  • Robotics & Simulation

    Embodied agents and the data pipelines behind them.
    • Vision-language-action models
    • MuJoCo
    • LeRobot
    • LIBERO
    • Reachy Mini SDK
    • 6-DOF quadrotor dynamics
    • 3D Gaussian Splatting
    • RGB/depth perception
    • Synthetic trajectory generation

    Where it shows up — GS-DroneGym on PyPI; Reachy Mini Talkie; IIT Mandi VLA research.

  • Languages

    Python by default. Rust and Zig when the runtime matters.
    • Python
    • Rust
    • Go
    • Zig
    • TypeScript / JavaScript
    • C
    • C++
    • Java
    • Swift
    • SQL
    • Bash
    • Julia
    • Haskell

    Where it shows up — LEMMA and Sieve in Rust, heapwatch in Zig, VERITAS in C++.

  • Backend & Data

    The part that has to stay up.
    • FastAPI
    • Flask
    • REST / WebSockets
    • Async jobs
    • PostgreSQL
    • Redis
    • MongoDB
    • Docker
    • Linux
    • CI/CD
    • Parquet
    • Hugging Face Datasets
    • Dataset versioning

    Where it shows up — UPAI Platform — 100+ models, PostgreSQL state, rate limiting, async jobs.

  • Security & Reliability

    Where the Cyber Sentinel win came from.
    • OAuth2 / OIDC
    • JWT
    • RBAC
    • OWASP Top 10
    • STRIDE threat modelling
    • Secrets handling
    • Input validation
    • Observability
    • Profiling
    • Incident response

    Where it shows up — Cyber Sentinel winner, IIT Madras; Sieve; Redauth.