Edra Labs AI Engineer · Jul 2024 to Sep 2026
Work Companion, an agentic LLM product
- Built the backend and agent orchestration for a product used by 1K to 10K people daily. Each turn runs episode routing, context assembly, an LLM call and token streaming over Socket.IO in under 300 ms (FastAPI, PostgreSQL, Redis).
- Designed its episode-based multi-agent system: each conversation episode gets its own agents and the context they need (user state, memory, prior turns), assembled on every turn.
Content Platform, LLM data extraction
- Cut new-category launch time from 6 months to 2 months and manual effort by 95% with an LLM platform that turns webpages, PDFs, images and video into structured data (RAG, semantic chunking, vector embeddings). Rolled out across 8 content categories.
- Led a 2-engineer team re-architecting the frontend layer into a modular Express backend: release cycles dropped 80%, and bulk async throughput rose 95% with AWS DMS replication and multi-queue Celery.
- Set up an isolated pre-prod environment with automated sync and one-click schema-aware deploys, which cut production issues by 70%.
Mineral Exploration, geophysical inversion on GPUs
- Owned the GPU and inference stack of a production inverse-modelling platform that identifies drill locations: CUDA/C++ forward models, Triton and CuPy kernels, Prefect workflows, H200 and B200 GPUs on Modal.
- Cut per-drillhole latency 3.9x (104 s to 27 s) on the same four A100s by sharding the reduction across GPUs over NCCL and fusing kernels; Bayesian ranking later reached 20 s per drillhole (43x faster).
- Built a diffusion-model approach to subsurface inversion: a UNet prior on synthetic 3D geology, conditioned on gravity and magnetic surveys with Diffusion Posterior Sampling to return several plausible models.
- Built the production inversion and falsification stack: MPPCA and MDS across 100K+ Monte-Carlo realisations with a fused CUDA falsification module at about 1 s per iteration.
- Led the satellite-spectral pipeline that narrowed exploration from continental scale to 10 km zones at 90% industry-grade accuracy, with weekly consultancy from IIT Bombay.