hey, i'm

directory

intro

• currently building paprika — runtime governance for ai agents. trace capture, policy enforcement, deterministic replay. deployed to 25+ b2b engineering teams.
• prev. swe intern @ Railinc (summer '24 + '25). react + java/spring boot on the RIGIS platform. etl pipeline processing 100K+ rail stations.
• studying cs @ berkeley (on leave)
• prev. MLE @ flywheel — the content engine for b2b SaaS teams. architected E2E MCP used by 10+ clients, wired voice, context, api keys, flags, and usage logging across 3 backend services.
• contributor to BFCL — the berkeley function calling leaderboard as part of RISE lab. drafted bfcl v5, a multimodal benchmark extension for agents → designed 9 perception-to-tool-call failure tests across audio/vision. cited by openai, nvidia, anthropic teams.
• researcher @ uc berkeley (eecs), satellite imagery + ML to detect crypto mining development in el salvador (Potts Lab)

projects

• routeX
  sub-millisecond routing engine for real-time navigation. uses contraction hierarchies + xgboost traffic predictions to find optimal routes across 280K+ road segments in the bay area.
  stack: c++ • postgis • xgboost • osm
• swandb
  point-in-time correct feature store with sub-8ms online serving. time-aware joins that actually work, no model training surprises.
  stack: duckdb • redis • parquet
• chronicle
  llm inference server with 32% throughput improvement over huggingface baseline. micro-batching + kv-cache reuse for sub-200ms p95 at 100 concurrent requests.
  stack: python • cuda • mistral-7b

open source

• vLLM / PyTorch
  merged performance optimization + distributed benchmarking PRs upstream.
• geofilter
  numba jit point-in-polygon engine. 10M points in 0.58s. wrote it because existing spatial filters were too slow and i was impatient.
• reqbench
  load-testing cli sustaining 4.8K req/sec at 500 concurrency with p95/error ci gating.

lore

• pretty good at geography trivia (top 1% sporcle, jetpunk player)
• i run buzzer based comps @ iac
• i've urbexed every part of the bay and will travel anywhere for a skyline or sunset pic
• played trevor rainbolt in ranked geoguessr. got absolutely cooked tho
• played carnatic violin for 10 years
• huge chelsea fan + chicago sports
• been to 30+ countries and my dream is to visit every single one
• have listened to 500 rap/rnb albums and will give better takes than anyone on the new releases

currently

• @ berkeley to scale and build wherever and whatever i can
• still thinking about that caleb williams throw…
• saving money for asap rocky, don toliver, and world cup games
• will bench 225 by the end of the year
• about halfway done with the blind 75!
• thailand summer 2026
• stay tuned for record labs. building in music, my passion.

some other things

• coursework: data structures, algorithms, discrete math, structure of computer programs, a.i., machine learning, networking + internet protocols, database systems, computer vision, data science
• languages: python, c++, typescript, java, sql
• ml + systems: FSDP, AMP, CUDA, KV-cache, distributed fine-tuning, eval pipelines, quantization
• frameworks: react, fastapi, spring boot, node.js, duckdb, langgraph, pytorch, xgboost
• infra: aws, docker, redis, postgresql, postgis, gdal, arcgis pro, linux, ci/cd
• interests: cities + public transportation, collecting old maps, tamil culture, soccer jerseys, geoguessr, chicago sports, all sorts of music
• less frequently: south indian movies, home decor, sketching, online chess, sparkling water

other things

• roast my spotify playlists
• i used to watch anime a bit
• trying to learn to dj and cook (or both at the same time)
• will run a marathon this year
• decent at sketching
• trying to attend more hackathons