Skip to content

Résumé

Atul Saxena

Research Engineer · RL Systems · Interpretability · ML Infrastructure

atul@blackdrome.tech · India · Remote

PDF

//Profile

I work on reinforcement-learning environments, mechanistic interpretability, and the infrastructure that holds both together.

Right now that means building production RL environments at Snorkel AI — task design, reward and verifier logic, and figuring out whether a failure came from the model, the environment, the grader, or the spec. Before that I led a pod at Bespoke Labs on agent evaluation and data curation for frontier-model work.

//Experience

  • RL Production Environment Engineer · Snorkel AI

    Jul 2026 – Present

    • Own task design, state transitions, tool execution, and reward/verifier logic for environments used in model training and evaluation.
    • Diagnose whether a failure came from the model, the environment, the evaluator, or the execution layer — then turn the discovered shortcut or regression into durable test coverage.
    • Validate environment quality and reproducibility so benchmarks expose genuine capability gaps rather than superficial pass/fail behaviour.

  • Founder — RL Infrastructure · Blackdrome AI Labs

    2026 – Present

    • Published three research artifacts with DOIs: a GPT-2 interpretability study, LEMMA (neuro-symbolic maths), and Noma (an open decision model).
    • Shipped Noma with open weights on Hugging Face under MPL-2.0, a published PyPI package, and ~16 ms end-to-end latency on a single GPU.
    • Maintains LEMMA, a Rust neuro-symbolic mathematics platform with 572 transformation rules and a released policy checkpoint.

  • Product & AI Systems (Contract) · Chalkbyte / Pyndra

    2026 – Present

    • Research user and business problems, decide what should be built, and translate partially defined needs into technical scope rather than waiting for a complete specification.
    • Iterate AI capabilities with engineering from real stakeholder feedback — applied LLM workflows, experimentation, evaluation and product integration.

  • Pod Lead — RL, Agent Evaluation & Data Curation · Bespoke Labs

    Mar 2026 – Jun 2026

    • Designed and reviewed evaluation harnesses and failure cases, focused on whether tasks and graders measured the intended capability rather than an exploitable shortcut.
    • Coordinated review and delivery across contributors, and contributed upstream to Bespoke Curator with normalized Mistral batch finish_reason handling plus regression coverage.

  • Researcher — Vision-Language-Action Robotics · IKSHMA Lab, IIT Mandi

    Mar 2026 – Aug 2026

    • Explored reproducible VLA experimentation and robotics data pipelines across language, visual observations, actions and simulated environments.
    • Work in this area produced GS-DroneGym, a published drone-learning environment with photorealistic Gaussian-Splatting rendering.

  • Co-Founder & ML Systems Engineer · PKBoost AI Labs

    2025 – Present

    • Build reproducible evaluation harnesses, dataset/versioning conventions and production-oriented model packaging.
    • Develop boosting/GBDT components with emphasis on low-latency inference, memory efficiency, drift-aware monitoring and deterministic rollback paths.

  • AI Mentor & Technical Lead · Edge Lab

    Jun 2025 – Sep 2025

    • Helped teams ship production-ready projects with clean interfaces and deployable pipelines.

  • Founder & Lead Developer · UPAI Platform

    May 2024 – Present

    • Built secure backend services with rate limiting, async jobs, retries, caching and PostgreSQL-backed state.
    • Built an LLM orchestrator and a benchmarking leaderboard for NVIDIA VIM.

  • Co-Founder & CTO · Boostra AI by GradeBoostHub

    Feb 2020 – May 2026

    • Database design, APIs, authentication, dashboards, production deployments, observability and incident response.
    • Operated Docker-based infrastructure with structured logging, monitoring, backups, RBAC, JWT/OAuth and hardened service boundaries.

//Research

  • Noma: An Open System One Decision Model from a Sliced Decoder and Trained Heads

    Atul Saxena, Divyanshi Sharma · October 2026 · 10.5281/zenodo.23186353 (opens in a new tab)

  • LEMMA: Learned Guidance for Evidence-Carrying Long-Horizon Symbolic Rewriting

    Atul Saxena, Pushp Kharat · September 2026 · 10.5281/zenodo.22842820 (opens in a new tab)

  • Investigating Trajectory-Conditioned Semantics and Dynamic Computational Pathways in GPT-2-Small

    Pushp Kharat, Atul Saxena · May 2026 · 10.5281/zenodo.20098563 (opens in a new tab)

  • From Paper to Working System: Reproducing AI Research When the Weights Aren't There

    Accepted 40-minute technical talk · Opportunity Open Source Conference 4.0, IIIT Allahabad

//Selected projects

  • GS-DroneGymGymnasium · 3D Gaussian Splatting · gsplat · PyTorch

    A pip-installable drone environment that combines 6-DOF quadrotor dynamics with photorealistic Gaussian-Splatting rendering, and the data tooling to train and evaluate VLA policies on it.

  • AXONTransformerLens · SAELens · FastAPI · Three.js

    Hooks GPT-2’s residual stream at layer 8, runs each token’s hidden state through a sparse autoencoder, and streams the top-K firing features into a live 3D force graph in the browser.

  • BrainPatchPyTorch · Transformers · vLLM · llama.cpp

    A portable file format and runtime for reversible activation-space interventions. No fine-tuning, no weight modification, no prompt injection, no hosted service.

  • LEMMARust · MCTS · Beam search · Ratatui

    An evidence-aware neuro-symbolic mathematics platform in Rust. It searches a stable registry of rules, records every transformation it applies, and returns the result with a precise verification status.

  • NomaPyTorch · Open weights · FastAPI · Calibration

    An open decision model for agents and pipelines. Give it a state and typed questions; get back calibrated probabilities, an abstain signal and an uncertainty measure. It never writes text.

//Open source

10 merged pull requests upstream, including:

//Skills

RL & Agents
RL / RLVR workflows · Reward & verifier design · Task generation · Long-horizon environments · Agent evaluation · Trajectory analysis · Failure-mode analysis · Gymnasium · Regression coverage for exploits
Interpretability
Mechanistic interpretability · TransformerLens · SAELens · Sparse autoencoders · Activation patching · Causal interventions · Probing classifiers · Ablations · Representation engineering · Activation steering
ML Systems
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
Robotics & Simulation
Vision-language-action models · MuJoCo · LeRobot · LIBERO · Reachy Mini SDK · 6-DOF quadrotor dynamics · 3D Gaussian Splatting · RGB/depth perception · Synthetic trajectory generation
Languages
Python · Rust · Go · Zig · TypeScript / JavaScript · C · C++ · Java · Swift · SQL · Bash · Julia · Haskell
Backend & Data
FastAPI · Flask · REST / WebSockets · Async jobs · PostgreSQL · Redis · MongoDB · Docker · Linux · CI/CD · Parquet · Hugging Face Datasets · Dataset versioning
Security & Reliability
OAuth2 / OIDC · JWT · RBAC · OWASP Top 10 · STRIDE threat modelling · Secrets handling · Input validation · Observability · Profiling · Incident response

//Education

B.S. in Data Science and Applications — Indian Institute of Technology Madras · Ongoing

Techkriti Hackathon Finalist · Cyber Sentinel Winner · Top 250, Shipped Pre-Accelerator · Live industry research project for HMC Group (Hero MotoCorp)

Professional Diploma in Agile & Project Management — MTF Institute, Portugal